the community origins of private enterprise in...
TRANSCRIPT
The Community Origins of Private Enterprise in China∗
Ruochen Dai† Dilip Mookherjee‡ Kaivan Munshi§ Xiaobo Zhang¶
February 28, 2019
Abstract
This paper identifies and quantifies the role played by birth-county-based community networks in thegrowth of private enterprise in China. The starting point for the analysis is the observation that populationdensity is positively associated with local social interactions, social homogeneity, and enforceable trust incounties (but not cities). This motivates a model of network-based spillovers that predicts how the dynamicsof firm entry, concentration, and firm size vary with birth county population density. The predictionsof the model are validated over the 1990-2009 period with administrative data covering the universe ofregistered firms. Competing non-network-based explanations can explain some, but not all of the results.We subsequently estimate the structural parameters of the model and conduct counter-factual simulations,which indicate that entry and capital stock over the 1995-2004 period would have been 40% lower withoutcommunity networks. Additional counter-factual simulations shed light on misallocation and industrialpolicy.
Keywords. Community Networks. Enforceable Trust. Entrepreneurship. Misallocation. Informal Institu-tions. Growth and Development.JEL. J12. J16. D31. I3.
∗We are grateful to Emily Breza, Matt Elliot, Debraj Ray, Daniel Xu, and numerous seminar and conference participants fortheir constructive comments. Research support from the Economic Development Initiative (EDI)and Cambridge-INET is gratefullyacknowledged. We are responsible for any errors that may remain.†Peking University [email protected]‡Boston University [email protected]§University of Cambridge [email protected]¶Peking University [email protected]
1 Introduction
China has witnessed the same degree of industrialization in three decades as Europe did in two centuries
(Summers, 2007). This economic transformation began in the early 1980’s with the establishment of township-
village enterprises (TVE’s) and accelerated with the entry of private firms in the 1990’s. Starting with almost
no private firms in 1990, there were 15 million registered private firms in 2014, accounting for over 90% of all
registered firms and 60% of aggregate industrial production. The surge in the number of private firms has had
major macroeconomic consequences. China is the largest exporter in the world today and, depending on how
the accounting is done, the world’s largest or second-largest economy (Wu, 2016).
What is perhaps most striking about the growth of private enterprise in China is that it occurred without
the preconditions generally believed to be necessary for market-based development; i.e. without effective
legal systems or well functioning financial institutions (Allen et al., 2005). While the government played an
important role in China’s economic transformation by providing infrastructure and credit (Long and Zhang,
2011; Wu, 2016), it has been argued that informal mechanisms based on reputation and trust must have been
at work to allow millions of entrepreneurs, most of whom were born in rural areas, to establish and grow
their businesses (Peng, 2004; Allen et al., 2005; Song et al., 2011; Greif and Tabellini, 2017; Zhang, 2017).
Case studies of production clusters; e.g. Fleisher et al. (2010) and Nee and Opper (2012) indicate that long-
established relationships among relatives and neighbors (from the rural origin) substitute for legally enforced
contracts between firms. These case studies document a high degree of mutual reliance of firms within the
cluster for exchange of intermediate goods, information, marketing connections and finance. Our analysis,
which utilizes comprehensive data covering the universe of registered firms over many years, advances this line
of research by identifying and quantifying the role played by informal community-based business networks in
the growth of private enterprise in China.
Our analysis proceeds in four steps. First, we argue that population density in rural counties, which is
mechanically correlated with spatial proximity, is positively associated with social connectedness. Spatial
proximity results in more frequent social interactions that, under plausible assumptions on the matching pro-
cess, give rise to more inter-connected social networks (Coleman, 1988; Jackson et al., 2012). Inter-connected
networks sustain greater economic cooperation via norms based on community enforcement (Greif, 1993, 1994;
Greif and Tabellini, 2017). The nature of trust and cooperation we focus on is distinct from from the general-
ized trust that has received much attention in the rapidly growing economics literature on culture (Alesina and
Giuliano, 2015). The domain of the former is restricted to local residents rather than the general population,
and is sustained via external enforcement rather than by internalized cultural values. We provide evidence
linking population density and localized trust with data from the population census and the China Family
Panel Survey (CFPS) which covers a nationally representative sample of households: the frequency of local
social interactions and trust in neighbors are both increasing in county population density, after controlling
for total population, occupation structure and average education. This result holds in counties, but not cities.
Moreover, even within counties, it applies only to trust in local residents rather than outsiders.
The argument that local cooperation is increasing in population density is, however, predicated on the
assumption that the local population is socially homogeneous, or that the degree of homogeneity is either
1
independent or increasing in population density. While there is a well established sociological literature that
describes the role of the hometown or birth county in supporting economic activity in China; e.g. Honig (1992),
Goodman (1995), it has also been claimed that the clan, a traditional kinship unit within the county, has
renewed its relevance in the post-collectivist era (Peng, 2004; Greif and Tabellini, 2017). We provide evidence,
using the CFPS and the population census, that clan concentration within counties (using surnames to identify
clans, as in Peng (2004)) is positively correlated with population density. In contrast, social homogeneity is
found to be decreasing in population density in cities, which may explain why local social interactions and
localized trust are not associated with urban population densities. The second step in our analysis, which links
population density to entrepreneurship, thus focuses on county-born businessmen. This is an important group
to study because their firms account for two-thirds of all registered private firms (and a comparable share of
total registered capital) in China. The majority of our county-born entrepreneurs establish their firms outside
their birth counties, often far away. Our key assumption, which is consistent with the literature on temporary
migration; e.g. Morten (2019) is that these entrepreneurs remain connected to, and can be sanctioned by, their
communities. If this assumption is satisfied, business networks drawn from higher population density counties
will support higher levels of mutual help among their members.
The canonical model of community-based entrepreneurship that we develop to validate the preceding
hypothesis features successive cohorts of agents that make a choice between a traditional occupation (such
as farming or wage labor) and becoming a private entrepreneur. Individual abilities are drawn from an i.i.d.
process and affect returns to both occupations; the payoff from entrepreneurship depends additionally on
what we refer to as community TFP (CTFP); i.e. the contribution of the network (defined by the set of
incumbent entrepreneurs from a common birth county) via mutual help. Help provided by different network
members is mutually complementary, which implies, in turn, that there are increasing returns to network size.
Moreover, the level of help (for given network size) is increasing in birth county population density, a measure
of ‘network quality’, for reasons discussed above. The model generates predictions for the dynamics of entry
across successive cohorts and the consequent evolution of network size, as a function of initial entry, population
density at the origin, and time. It features a network ‘multiplier’ where higher initial entry generates greater
entry flows in later cohorts, with the multiplier increasing in population density and in time.
These predictions are tested with administrative data obtained from the State Administration of Industry
and Commerce (SAIC) covering the universe of registered firms in China. The following information is available
for each private firm: establishment date, sector, location, registered capital, and a list of major shareholders
and managers, with their citizenship ID. The county of birth can be extracted from the citizenship ID and
the firm’s legal representative is designated as the “entrepreneur” for our analysis. Since the first wave of
private entry in China commenced in the early 1990’s, we estimate the effect of initial entry in 1990-1994
on subsequent entry, separately in 2000-2004 and 2005-2009, at the level of the birth county-sector-location.
All counties and cities (urban districts) in the country are included in the set of locations in the benchmark
specification, although we verify that the results are robust to excluding the birth county itself. Consistent
with a dynamic network multiplier effect, we find that one additional entrant from the birth county in a
particular sector-location in the initial 1990-1994 period is associated with seven additional entrants over the
2000-2004 period and nine additional entrants over the 2005-2009 period.
2
While these results highlight the inter-linkage between firms from the same birth county, non-network
explanations for the inter-temporal correlation are also available. A stronger test that networks are active is
that the initial entry effect should be increasing in birth county population density, regardless of where firms
are located, and this is indeed what we find. We also find that conditional on the number of initial entrants
from the birth county, the total number of initial entrants in a sector-location (aggregating across all origins)
does not predict the number of subsequent entrants. This is consistent with the hypothesis that the birth
county is the domain within which business networks are organized in China, rather than destination sectors
or locations. The absence of a (negative) cross-community effect also goes against the view that the networks
are simply competing with each other for subsidized credit from the local government, as in Bai et al. (2019),
although our framework does not rule out the presence of such rent seeking.
We supplement these results by testing the hypothesis that the primary locus of informal cooperation
within a county is the clan. Using birth county and surname to measure clan affiliation, and implementing the
same test as above, we find that network spillovers are concentrated within the clan, although greater initial
entry by non-clan members from the birth county also has a positive and significant (albeit smaller) effect on
subsequent entry. Assuming that the spillovers are restricted to the clan, the model predicts in addition that
entering entrepreneurs will be increasingly concentrated in particular clans, with a steeper increase in clan
concentration over time for entrepreneurs from higher population density birth counties. The data support
each of these predictions, highlighting the sociological roots of private enterprise in China.
The third step of our analysis augments the model of community-based entrepreneurship to incorporate
sector-location choice and capital investment. Although the augmented model is set up to match the basic
features of the canonical model, there are now two sources of network-based spillovers: (i) post-entry mutual
help among networks members, as before, and (ii) a pre-entry referral process which increasingly channels
entering firms from a given origin into an initially favored destination (a term we use to denote either sectors
or locations). The interaction between the two types of spillovers generates dynamic increasing returns to
network size in any given destination, resulting in increased sectoral and spatial (within sector) concentration
over time. Concentration is also rising in origin population density at each point in time, with a slope that
increases over time at early stages of the industrialization process.
With regard to capital investment, the network-based spillovers that raise CTFP over time have two
conflicting effects on the initial size of the marginal entrant’s firm: the direct effect, for a given level of ability,
is to increase firm size by raising the firm’s TFP, but an increase in CTFP also lowers the ability threshold for
entry into entrepreneurship and this negative selection works in the opposite direction to lower TFP. We show
that the latter effect dominates; the marginal entering firm from a higher population density birth county will
be unambiguously smaller, with this negative relationship growing stronger over time as networks get larger.
Under specific conditions on the model’s parameters, this result is shown to hold for average initial firm size
as well. This contrasts with the model’s predictions for the post-entry growth in firm size. This growth is
driven by changes in CTFP over time and is the same for all firms in a network at a given point in time,
regardless of their cohort or the ability of the entrepreneur. Because networks from higher population density
birth counties are growing faster, firms from those counties will start small but subsequently grow faster.
The predictions of the augmented model are tested over the 1990-2009 period with SAIC registration data,
3
supplemented with the industrial census and the SAIC inspection database for the analysis of firm growth.
Population, education, and the occupational structure (measured in the birth county in 1982) are included
in the estimating equations as controls. Population density could, nevertheless, be correlated with other
unobserved determinants of the outcomes of interest and, hence, we focus on the less easily explained interaction
between birth county population density and time; this is similar to a difference-in-difference analysis, except
that we are leveraging the dynamic implications of the model. Exploiting the fact that firms from many
birth counties are established in the same sector-location, we also include sector-time period and location-time
period effects in the estimating equations. This effectively controls for access to resources provided by local
governments, and for the destination-based productivity spillovers emphasized in the literatures on endogenous
growth (Romer, 1990, 1986; Segerstrom et al., 1990; Aghion and Howitt, 1992; Jones, 1995; Segerstrom, 1998)
and agglomeration (Ciccone and Hall, 1996; Ellison and Glaeser, 1997; Au and Henderson, 2006; Combes
et al., 2012). The model generates dynamic predictions for the relationship between birth county population
density and a rich set of outcomes – firm entry, sectoral and spatial concentration, firm size – and the data
match each of them. Moreover, the relationships get stronger when sector and location effects are added to the
estimating equation. This indicates that entrepreneurs from higher population density birth counties are, if
anything, selecting into less favorable sectors and locations. As a final check, we consider a comprehensive set
of potential non-network explanations for our results, which are obtained by systematically relaxing different
assumptions of the model. We show that while these explanations can explain some of the results, there is no
other explanation that can account for all of them simultaneously.
There are many anecdotal examples of the role played by communities in supporting the business activities
of their members, in China and elsewhere (see Munshi (2014) for a survey). There have also been some
attempts; e.g. Banerjee and Munshi (2004), Munshi (2011), to estimate the impact of community networks on
selection into entrepreneurship and firm outcomes in specific industries and locations. However, our analysis
is the first that we are aware of to document the role played by communities in the evolution of private
enterprise in an entire economy. Having tested and validated the network-based model, the final step in the
analysis thus seeks to quantify the impact of these networks on aggregate firm entry and capital stocks by
estimating the structural parameters of the model and then conducting counter-factual simulations. Although
the model is extremely parsimonious, it does a good job of matching entry and initial capital within sectors,
across the range of birth county population densities, within the sample period for the structural estimation
(1995-2004) and out of sample. This increases our confidence in the results of a counter-factual experiment,
which estimates that entry from county origins would have declined by as much as 64% over the 1995-2004
period (with a comparable decline in the total capital stock) had the networks not been active. As entry and
capital stocks from county origins accounted for approximately two-thirds of all entry and capital in China,
this amounts to an impact of approximately 40% for the entire country. Given the dynamic increasing returns
that are generated by the networks, the long-term consequences of their absence would have been even more
substantial.
Two stylized facts motivate a large and growing literature in macro-development: (i) the variation in
marginal productivity and, hence, firm size within narrow sectors is especially wide in developing economies,
and (ii) firms in those economies are small (Peters, 2016). Although a number of mechanisms can explain
4
these facts; e.g. Caunedo (2016), Asker et al. (2014), Akcigit et al. (2016), Haltiwanger et al. (2018), perhaps
the simplest is based on a model with mark-ups in output prices and wedges in factor prices (Restuccia and
Rogerson, 2008; Hsieh and Klenow, 2009, 2014; Peters, 2016). There are no price distortions in our model.
Our objective is to demonstrate that wide dispersion in firm size and productivity can be a consequence of
community-based inter-firm spillovers, rather than inefficient taxes or regulations. The results of a second
counter-factual policy experiment, which provides a temporary credit subsidy to entering firms, suggests that
optimal second-best policies should target these subsidies to more socially connected communities, as a way of
exploiting the resulting network spillovers from increased entry. This would increase increasing dispersion and
the proportion of small firms even further. More generally, we would not want to infer that one developing
economy is less efficient than another developing economy because it has smaller firms or greater dispersion in
firm size. Indeed, these characteristics may be symptomatic of a more dynamic economy in which underlying
community networks are responding more effectively to market frictions. Although this insight, and the
message that policies aimed at raising growth and efficiency should incorporate intra-community spillovers over
and above individual ability, would apply to all economies where community networks are active, the important
qualifier is that the magnitude of the estimated network effects are specific to China and would not necessarily
extend to other countries with different social structures and economic infrastructure. Moreover, policies that
attempt to exploit these network effects will have complex distributional consequences; in particular, policies
that target more connected communities are likely to exacerbate inter-community inequality, while promoting
intra-community equity.
2 Institutional Setting
2.1 Private Enterprise in China
The State Administration of Industry and Commerce (SAIC) database that we utilize for much of the empirical
analysis comprises the universe of registered firms in China, regardless of their size, from 1980 onwards. These
firms are classified as township-village enterprises (TVE’s), state owned enterprises (SOE’s), foreign owned
firms, and private (domestically owned) firms. Our analysis focuses on the private firms. The SAIC database
lists the major shareholders and managers, with their citizenship ID, in each registered private firm. New
firms enter the database each year, while a fraction of incumbents exit. We can thus trace the initial growth
phase of private enterprise in China in its entirety; starting with a relatively small number of private firms
in 1990, there were close to 15 million registered private firms in 2014 (and 7.3 million private firms in 2009,
which will be the end point of our statistical analysis).1 As documented in Figure 1a, private firms accounted
for approximately 10% of all firms in the early 1990’s. By 2014 they accounted for over 90% of all firms.2
1In contrast, previous analyses of firms in China have relied on a publicly available database of manufacturing firms with salesabove a threshold level (5 million Yuan) over the shorter 1998-2008 period; e.g. Hsieh and Klenow (2009), Song et al. (2011),Brandt et al. (2012), Aghion et al. (2015). The above-scale firms account for less than 15% of all private firms in the registrationdatabase in 2008.
2Appendix Figure D.1 reports the number of private firms, TVE’s, and SOE’s, in each year over the 1990-2014 period. As canbe seen, the number of private firms is an order of magnitude larger than the number of TVE’s and SOE’s by the early 2000’s,with the gap continuing to grow steeply thereafter. The ownership type assigned to a given firm in the figure is based on its finaltype (in the event that it was privatized). However, just 5.5% of registered private firms report any change in their ownershipstructure, which includes changes in their ownership type. Given the dominance of the private firms in terms of their numbers,their growth evidently could not have been generated simply by privatizing existing TVE’s and SOE’s.
5
Figure 1b reports the share of total registered capital, by firm-type, over the 1980-2014 period.3 As with their
numbers, the share of registered capital held by private firms grew steeply from the early 1990’s onwards and
by 2014 they held 60% of total registered capital in the economy.
Figure 1. Distribution of Firms, by Type
0.2
.4.6
.81
Fra
ctio
n of
firm
s
1980 1985 1990 1995 2000 2005 2010 2015Year
Foreign TVE SOE Private
(a) Number of Firms0
.2.4
.6.8
1F
ract
ion
of r
egis
tere
d ca
pita
l
1980 1985 1990 1995 2000 2005 2010 2015Year
Foreign TVE SOE Private
(b) Registered Capital
Source: SAIC registration database.
The preceding facts suggest that private firms played an important role in China’s rapid growth over the
past decades. This raises the natural question about the underlying cause of this surge in private enterprise. It
is generally believed that governments at the local (county), provincial, and central level played a substantial
role in China’s economic transformation. Local governments provided the infrastructure to support production
clusters located throughout the country, which are a distinctive feature of the Chinese economy (Long and
Zhang, 2011). Provincial governments and the central government supported firms by giving them subsidized
credit and by aggressively promoting exports (Wu, 2016). Our interest, however, is in the role played by
informal community-based institutions in supporting private entrepreneurship. Allen et al. (2005) argue that
reputation and relationships must have substituted for missing financial institutions for China to grow so
rapidly. Song et al. (2011) explain China’s unique growth path as consequence of the fact that more productive
private firms had to rely on self-financing in the absence of low cost formal finance. Case studies of Chinese
production clusters; e.g. Huang et al. (2008), Ruan and Zhang (2009), and Fleisher et al. (2010) consistently
find that the impetus for their formation came from within, with groups of entrepreneurs setting up firms with
little external support. The involvement of local governments is found to come later, through the provision of
infrastructure such as roads, markets, and quality control.
What specific institutional arrangements allowed these early entrants to come together and become private
entrepreneurs? There are many accounts of the role played by social networks or guanxi in facilitating China’s
historically unprecedented rural-urban labor migration over the past decades; e.g. Zhao (2003), Zhang and Li
(2003), Hu (2008). These accounts describe how migrant networks are organized around the rural hometown,
3The initial registered capital represents the total amount paid up by the shareholders. This amount is deposited with theSAIC, and can be used to pay the firm’s operating expenses before it becomes cash flow positive. Access to bank credit is alsodependent on the firm’s registered capital, which is why firms will often choose to increase their registered capital over time.
6
complementing an older anthropological literature that takes the position that ethnicity in China is defined
by the native place; e.g. Honig (1992, 1996), Goodman (1995).4 If the sending county is the domain around
which migrant labor networks are organized, then it is plausibly the natural domain around which business
networks supporting county-born entrepreneurs are organized.5
Counties in China are divided into villages, which consist, in turn, of one or more clans or lineages (Peng,
2004; Tsai, 2007). These clans historically supported the business activities of their members, who were bound
together by mutual moral obligations. It has been argued that this role has re-emerged in the post-collectivist
era (Peng, 2004; Zhang, 2017; Greif and Tabellini, 2017). We remain agnostic about the boundary of the
social unit from which business networks are drawn in this paper; i.e. whether it is the county or the clan.
As discussed below, the county characteristic that we use as the source of forcing variation in the empirical
analysis would apply to the county as a whole and to all clans within the county, and thus our results would
apply in either case.
2.2 Population Density and Trust in Chinese Counties
The point of departure for our analysis is the assumption that social connectedness in Chinese counties is
increasing in population density. The underlying idea is that greater spatial proximity raises the frequency
of social interactions and facilitates communication with neighbors. This, in turn, helps sustain higher levels
of mutual cooperation, supported by the threat of social sanctions, as argued in early papers on social norms
and community enforcement (Greif, 1993, 1994; Kandori, 1992; Ellison, 1994). To make this argument more
precise, consider a random graph model in which the probability that an individual is connected to any other
individual in a local population is equal to γ, which is rising in spatial proximity. A higher γ directly raises the
degree of the social network (the number of links per capita), and indirectly also network inter-connectedness
i.e. the probability that friends of friends are linked, and so on. For example, the rate of triadic closure –
the probability that any three individuals are directly linked – is increasing in γ. Coleman (1988) argues that
network closure is a necessary condition for economic cooperation, enforced by social sanctions. Jackson et al.
(2012) make a similar argument based on a related network property, which they refer to as “support.” These
results do not necessarily rely on the random matching assumption, and are also likely to hold if the matching
process exhibits homophily.6
The preceding discussion implicitly assumes that the county is socially homogeneous and comprises a single
community. Matters are more complex when the county population is fragmented into smaller communities.
For example, suppose that each county is composed of a number of clans j = 1, . . ., where cooperation
4Migrants from the same rural origin move to the city in groups and most migrants end up living and working with laoxiangor “native-place fellows” (Fang, 1997; Ma and Xiang, 1998; Zhang and Xie, 2013). In Chinese cities, migrant-peasant enclavesare often named after a sending province, but as Ma and Xiang (1998) note, this nomenclature is misleading because the enclavetypically consists of peasants from a single county or two neighboring counties.
5A similar argument has been made in past research in India, where the endogamous caste or jati is the common domainaround which networks supporting rural-urban migration, business, and other functions are organized (Munshi and Rosenzweig,2006, 2016; Munshi, 2011).
6Jackson and Rogers (2007) describe a model of matching in which some links are formed randomly, whereas other links areformed strategically. The latter process results in clustering or homophily, with individuals having higher degree (more links)more likely to be matched to each other. While adding a strategic element to the matching process will generally increase theinter-connectedness of the network, the relationship between population density (γ) and inter-connectedness now becomes morecomplex. Nevertheless, we continue to expect this relationship to be positive by a limit argument; as γ goes to zero, no one isconnected and the rate of triadic closure goes to zero. As γ goes to one, everyone is connected, and the rate of triadic closure goesto one.
7
is sustained only within clans, not across clans. Let ζj denote the demographic weight of clan j, so clan
composition of the county is represented by the vector ζ ≡ {ζj}j . The likelihood of sanctions being applied by
clan members is increasing in the population density p (which determines the frequency of interactions within
every clan in the county), while the magnitude of the sanction is rising in the relative size of the clan (e.g.,
a larger clan will correspond to a higher proportion of neighbors that impose sanctions on any deviant). The
total expected sanction and, hence, the level of cooperation in the clan is the product of two functions, ν1(p)
which is rising in population density p, and ν2(ζ) which is rising in the relative size of the clan. Aggregating
across clans, average cooperation in the county equals ν1(p)∑
l ζlν2(ζl). If clan composition is independent
of p, then again a rise in population density is associated with higher average cooperation. However, if clan
composition is correlated with p, the effect of a higher p is complicated as it depends on the nature of this
correlation, as well as the curvature of the function ν2(ζ).7 Nevertheless, if clan concentration is positively
associated with p, and greater concentration (which implies greater social homogeneity) is associated with an
increase in cooperation — as frequently observed in practice8 — then higher population density will continue
to be associated with higher average cooperation in the county.
The assumption that social homogeneity is increasing in population density is unlikely to apply to cities.
Most urban residents in developing countries are recent arrivals, typically from diverse origins. Greater popula-
tion density in an urban area may thus reflect greater diversity of origins, rather than higher intra-community
density. The social enforcement that can be supported within long-established clans or among local residents
in counties (who have been living together for generations) thus will not be necessarily sustainable in urban
neighborhoods.
The preceding discussion indicates that cooperation is increasing in population density in a socially hetero-
geneous population if both the frequency of local social interactions and social homogeneity are increasing in
population density. We use the China Family Panel Survey (CFPS), which covers a nationally representative
sample of households, to provide empirical support for each of these relationships, in counties but not in cities.9
Table 1 reports the relationship between population density and local social interactions, separately in counties
and cities, based on the 2010 round of the CFPS. We divide the different types of social interactions reported in
the family module of the CFPS into planned interactions; i.e. group entertainment, visits to neighbors’ homes,
and dining together, and unplanned interactions, which are defined as meetings or conversations without other
background activities.10 Population density in the counties and cities covered by the CFPS, and in all the
analysis that follows, is computed from the 1982 population census. This is before the first wave of privatiza-
tion in the early 1990’s and the accompanying rural-urban migration, which could have endogenously shifted
the population density in ways that are correlated with the outcomes of interest.11 Moreover, the estimating
7A mean preserving increase in spread of the distribution {ζj}j will increase (resp. decrease) average cooperation if ν2 is convex(resp. concave).
8Peng (2004) and Zhang (2017) show that entrepreneurship and employment in private enterprise in China are positivelyassociated with measures of clan concentration within villages or prefectures, after controlling for local area characteristics.
9There are approximately 2,000 counties and 250 prefecture-level and province-level cities (which are further divided into urbandistricts) in China.
10Group entertainment includes playing mahjong or cards, reading newspapers, listening to the radio, or watching TV withothers.
11Population density is measured in units of 10,000 people per square km. The threshold density in our analysis is set at 0.002;i.e. 20 people per square km. This excludes sparsely populated regions such as Western China, Inner Mongolia, and Tibet, whichare inhabited by ethnic minorities with a different culture than the majority Han Chinese.
8
equation in Table 1, and all the analyses in the paper that estimate the direct effect of population density,
include population, education, and occupational structure in the county or city to allow for the possibility that
these variables (which are correlated with population density) could be directly determining the outcomes.
Columns 1-2 and 4-5 report the relationship between population density and the frequency of each category
of local social interactions, per month, for respondents residing in counties and cities, respectively. The popu-
lation density coefficient is positive (and statistically significant with unplanned interactions as the dependent
variable) for county residents, whereas it is negative (and statistically insignificant) for city residents. Table 1,
Columns 3 and 6 provide additional support for this distinction between counties and cities by estimating the
relationship between population density and a different measure of local social interactions obtained from the
adult individual module of the CFPS. This measure is based on a question that ascertains who the respondent
has the most unplanned social interactions with: local residents, relatives, classmates, colleagues, or others.
We construct a binary variable, which indicates whether the respondent lists local residents as his most fre-
quent interaction partners. With this measure as the dependent variable, the population density coefficient is
positive and statistically significant for county residents (Column 3) and negative and statistically insignificant
for city residents (Column 6).
Table 1. Frequency of Local Social Interactions and Population Density
Respondent’s location: county city
Dependent variable: frequency of social in-teractions per monthwith local residents
indicates whethermost unplannedinteractions arewith localresidents
frequency of social in-teractions per monthwith local residents
indicates whethermost unplannedinteractions arewith localresidentsType of social
interactions: planned unplanned planned unplanned
(1) (2) (3) (4) (5) (6)
Population density 0.343 1.635** 0.022** -0.778* -0.573 -0.019(0.422) (0.789) (0.010) (0.398) (0.834) (0.012)
Mean of dependentvariable
3.907 15.54 0.197 3.434 13.56 0.124
Observations 8,830 8,830 20,669 3,255 3,255 7,088
Source: China Family Panel Survey (2010). Columns 1-2 and 4-5 based on family module and Columns 3 and 6 based on adult individualmodule.Planned interactions include group entertainment, visits to neighbors’ homes, and dining together. Unplanned interactions are one-on-onesocial meetings without other background activities.Population density is measured in units of 10,000 people per square km, and then converted to Z-score.Control variables include population, education and occupation distribution in the birth place.Population density, population, education and occupation distribution are computed from the 1982 population census.Standard errors clustered at the county or city level are reported in parentheses. * significant at 10%, ** at 5%, *** at 1%.
Table 2 provides an explanation for the observed differences in social interactions between counties and cities
based on their distinct social structures. We begin, in Columns 1 and 3, by establishing that population density
at the aggregate – city or county – level, obtained from the 1982 population census, is positively associated with
population density at the local – neighborhood or village – level, obtained from CFPS (2010). Next, we compare
the fraction of locally born residents in counties and cities in Columns 2 and 4, respectively. Based on the mean
of the dependent variable, these fractions are very different; while around 90% of county residents are born
there, this statistic is lower than 50% in cities. Moreover, the fraction of locally born residents is decreasing
in local population density in cities (Column 2) but not counties (Column 4), and this increasing social
9
heterogeneity might explain the weakly negative relationship between population density and social interactions
that we observe in the cities. In contrast, Column 5 indicates that social homogeneity, appropriately defined,
could even be increasing in population density in Chinese counties. If economic cooperation is organized within
the clan, then the appropriate measure of social homogeneity is clan concentration. Members of a clan will
share the same surname (Peng, 2004) and the CFPS (2010) provides the fraction of the village population that
has the most popular surname. We see in Table 2, Column 5 that local population density has a positive and
significant effect on this measure of social homogeneity. Clan affiliation is less relevant in the city, particularly
among the urban born, which is why the surname information is not collected for urban residents in the CFPS.
Table 2. Social Structure and Population Density
Respondent’s location: city county
Dependent variable:population densityat neighborhood
level
fraction ofresidents born
locally (%)
populationdensity at
village level
fraction ofresidents born
locally (%)
fraction of householdsfrom the largest clan
(%)
(1) (2) (3) (4) (5)
Population density atcity/county level 3.266*** – 2.077*** – –
(0.993) (0.548)Population density atneighborhood/village level
– -7.750*** – 1.567 18.460*
(2.591) (2.109) (9.608)
Mean of dependent variable 0.772 49.27 0.112 90.14 42.46Observations 124 124 264 264 264
Source: China Family Panel Survey (2010).Clan affiliation is measured by surname. Population density at neighborhood/village level is computed from the community module of theCFPS. Population density at city/county level is computed from the 1982 population census.Population density is measured in units of 10,000 people per square km.Robust standard errors are reported in parentheses. * significant at 10%, ** at 5%, *** at 1%.
Figure 2a nonparametrically estimates the relationship between the fraction of residents born locally and
population density, separately in counties and cities. The additional information provided by the figure is that
although the population density is greater in cities than in counties on average, there is substantial overlap in
the range. The gap in the fraction of residents born locally, between counties and cities, is retained even in
the region where the population densities overlap.
The preceding results suggest that trust or economic cooperation should be increasing in population density
in counties, but not cities. The adult individual module of the CFPS (2012) collected information on trust,
which proxies for economic cooperation. Trust is measured as an ordinal variable, taking values from 0 to
10.12 We see in Table 3, Columns 1 and 3 that trust in local residents is increasing in population density
for respondents residing in counties, but not in cities. Recall that the localized and enforceable trust that is
relevant for our analysis is distinct from the generalized trust that has received much attention in the culture
literature. The implicit assumption in our analysis is that generalized trust is uncorrelated with population
density in counties and in cities, and this is indeed what we observe in Columns 2 and 4, with population
12The question on trust is designed to match the well known and frequently used question on trust in the World Values Survey(WVS). The WVS measures trust as an ordinal variable that takes values from 1 to 4. The level of trust in the WVS is assessedfor the respondent’s family, in his neighborhood, among people that the respondent knows personally, among people he meets forthe first time, among people of another religion, and among people of another nationality. The CFPS measures trust in parents,neighbors, Americans, strangers, cadres, and doctors. Our analysis focuses on trust in neighbors or local residents and strangers;i.e. people the respondent meets for the first time (and who thus live outside the local area). These are the two categories thatoverlap with the WVS.
10
Figure 2. Fraction of Residents Born Locally, Trust and Population Density
40
50
60
70
80
90
Frac
tion
of R
esid
ents
Bor
n Lo
cally
(%)
0 .05 .1 .15 .2 .25
Population density
county city
(a) Fraction of Residents Born Locally
6
6.2
6.4
6.6
6.8
0 .05 .1 .15 .2 .25
Population density
county city
(b) Trust in Local Residents
Source: China Family Panel Survey. Fraction of residents born locally is calculated from the adult individual module of China FamilyPanel Survey (2010) Trust measures are obtained from the adult individual module of China Family Panel Survey (2012).
density having no impact on trust in outsiders (which plausibly measures generalized trust).
Table 3. Trust and Population Density
Dependent variable: trust in local residents trust in outsiders trust in local residents trust in outsiders
Respondent’s location: county city
(1) (2) (3) (4)
Population density 0.225*** -0.018 -0.073 -0.014(0.054) (0.079) (0.056) (0.064)
Mean of dependent variable 6.429 2.194 6.239 2.115Observations 20,047 20,047 6,236 6,236Source: Trust measures are obtained from the adult individual module of China Family Panel Survey (2012).Trust is an ordinal variable, and takes value from 0 to 10.Population density is measured in units of 10,000 people per square km, and then converted to Z-score.Control variables include population, education and occupation distribution in the county or city.Population is measured in millions and education is measured by the percent of the population that is literate.Occupation distribution is measured by the share of workers in agriculture and industry with services the excluded category.Population density, population, education and occupation distribution are computed from the 1982 population census.Standard errors clustered at the county or city level are reported in parentheses. * significant at 10%, ** at 5%, *** at 1%.
Figure 2b reports nonparametric estimates corresponding to Table 3, Columns 1-2, but without the ad-
ditional covariates. Localized trust is increasing in population density in counties, but not in cities. This
distinction is retained even in the region where the population densities overlap. Our analysis of community-
based entrepreneurship, which uses population density as the source of forcing variation, is thus restricted to
county-born entrepreneurs.13 Figure 3 switches back to the SAIC data, describing the growth in the total
number of private registered firms and in the number of firms owned by county-born entrepreneurs, respec-
tively, over the 1990-2014 period. We see that county-born entrepreneurs made up about two-thirds of all
entrepreneurs in China, with this ratio remaining quite stable over time. Firms owned by county-born en-
trepreneurs are just slightly smaller than the average registered firm (not reported). The contribution of these
13This does not imply that city-born entrepreneurs do not have access to networks; based on the evidence provided above,population density is just not a good indicator of social connectedness for them.
11
entrepreneurs, most of whom are first-generation businessmen, to Chinese economic growth has thus been
substantial. Our objective in the analysis that follows is to identify and quantify the role played by birth
county networks in supporting the explosive entry and subsequent growth of their firms.
Figure 3. Growth of Private Enterprise, by Birthplace of Entrepreneurs
0
5
10
15
Num
ber
of fi
rms(
mill
ion)
1990 1995 2000 2005 2010 2015
Year
All entrepreneurs County-born entrepreneurs
Source: SAIC registration database.
2.3 Population Density and Entrepreneurship: A Canonical Model
Entrepreneurs in developing countries often rely on other entrepreneurs operating in the same sector and
location for different forms of help. Chinese production clusters, for example, are characterized by a high degree
of specialization by firms in specific stages of production, extensive exchange of intermediate goods, and flexible
adjustment to workloads and product specifications in response to volatile market demand. Entrepreneurs also
rely on each other for connections to buyers and suppliers (who often provide credit) as well as for information
about new technologies and markets. Sustaining high productivity in this economic environment thus requires
considerable mutual help. This is difficult to generate via market transactions, due to the inherent problem of
verifying help sought and received, coupled with a weak legal environment. Cooperation is based instead on
community norms, backed by social ties among the entrepreneurs in question (Nee and Opper, 2012). The key
assumption in our analysis is that these social ties are based on the birth county, despite the fact that a majority
of county-born entrepreneurs establish their firms elsewhere. Migrant entrepreneurs typically have close family
members, such as aged parents, in the birth county, and visit it frequently. They (or close family members)
thus continue to interact socially in the birth county. It follows that cooperation in the sector-location where
firms are operating can be backed effectively by social sanctions imposed by the wider community in the birth
county. The discussion that follows derives the relationship between population density in the birth county,
network quality and, hence, the returns to entrepreneurship.
Let the size of the business network, which consists of entrepreneurs from the same origin who have already
entered a particular sector-location in period t be denoted by nt. The TFP of an entrepreneur with individual
ability ω in this sector-location equals f(ω)At where f is an increasing function of ω and At = A0(1 + h)nt−1 ,
12
with h denoting help provided by incumbent members of the network to one another. This specification
captures the idea that help provided by network members is mutually complementary, which implies, in turn,
that there are increasing returns with respect to the size of the network (for given per-member help). If the
help provided by one network member to another is observed by other network members and entrepreneurs
remain connected to their origin communities, then the maximal incentive compatible level of help (for given
network size) will be increasing in population density and social homogeneity in the birth county. Given that
social homogeneity, measured by clan concentration, is increasing in population density, it follows that the
equilibrium level of help, h(p), which reflects the ‘quality’ of the network, will be unambiguously increasing in
the population density of the birth county, regardless of whether the network is drawn from the clan or the
county as a whole. Hence At = A0(1 + h(p))nt−1 . Letting θ′(p) denote log(1 + h(p)), this reduces to
At = A0 exp(θ′(p)nt−1) (1)
We will use this expression for community TFP (CTFP thereafter) to derive testable predictions for business
choices and outcomes when networks are active.
But before we do so, it is useful to note the difference between this specification and others more commonly
used in the enodogenous growth and agglomeration literatures. In our formulation, networks are based on the
social origins of entrepreneurs rather than the destinations they select: θ′ reflects social connectedness in the
birth county and nt is the number of firms from that county operating in a given destination. In the standard
agglomeration model, the θ′ parameter would measure exogenous destination characteristics and nt would be
the total number of firms operating in that destination, irrespective of the social origin of their respective
entrepreneurs. Ciccone and Hall (1996), for instance, use the number of workers per square km as a proxy for
agglomeration effects in a given location. We will exploit this difference in the empirical analysis to distinguish
between birth county network effects and agglomeration effects.
To derive the implications of our formulation for the dynamics of entry, consider a given origin county
with social connectedness θ′(p), which is increasing in population density, p. There are equal sized cohorts
of new agents born at successive dates t = 1, 2... Each agent makes a once-and-for-all choice between a
traditional occupation (such as farming or wage labor) and becoming an entrepreneur in the single sector-
location that is available. Agents vary in individual ability ω, which is drawn independently from a log
uniform distribution on the unit interval. The returns to entering the traditional occupation for an agent of
ability ω is ωσ where 0 < σ < 1, while the return from entrepreneurship is Atω, where CTFP At given by (1)
depends on the incumbent stock nt−1, i.e., the total number of entrepreneurs from previous cohorts from the
same origin. The linear ability term in the production function ensures that there will be positive selection
on ability into entrepreneurship. We will see below that the linearity is derived endogenously, and not by
assumption, in the augmented model that includes capital in the production function. Higher ability agents
thus become entrepreneurs, while low ability agents remain in the traditional occupation, with the precise
threshold depending on the incumbent stock. The threshold ability, where returns to the occupations are
equalized, is seen to be
logωt = − 1
1− σ[logA0 + θ′(p)nt−1] (2)
We assume that the threshold lies in the interior of the unit interval. The model thus applies to the ‘early
13
industrialization’ phase, which seems empirically relevant for the period we consider in China.14 The fraction
of cohort t agents that enter is then given by
et = 1 +1
1− σ[logA0 + θ′(p)nt−1] (3)
which can be written as
et = B + θ(p)nt−1 (4)
with B ≡ 1 + 11−σ logA0 > 0 and θ(p) ≡ θ′(p) 1
1−σ .
Solving recursively, equation (4) yields the following expression for the evolution of entry flows as a function
of initial entry n0 and origin social connectedness, θ(p):
et = (B + n0)(1 + θ(p))t−1 (5)
It follows from equation (5) that entry in any period t is increasing in initial entry. Moreover, the initial
entry effect is increasing over time due to its compounding effect on future entry when networks are active.
These dynamic increasing returns to initial entry result in an increase in entry over time (this is easy to verify
because θ > 0), which is consistent with the explosive growth in the stock of firms that we observed in Figure
3. At any point in time, the initial entry effect is increasing in θ(p) and, hence, in origin population density,
p. Finally, the effect of the initial entry - population density interaction, which is positive, will grow stronger
over time, once again due to the compounding network effect.15
2.4 Testing the Predictions of the Canonical Model
Although the model assumes that there is a single entrepreneur in each firm, in practice most registered firms
consist of multiple shareholders. The SAIC database lists the major shareholders and managers, with their
citizenship ID, in each registered private firm. The first six digits of the citizenship ID reveal the birth county
of the individual.16 We designate the firm’s legal representative as the “entrepreneur” for the purpose of
the empirical analysis and his birth county thus applies to the firm as a whole. This individual is legally
responsible for the firm’s liabilities and typically plays a key role in its functioning; for example, 75% of legal
representatives are shareholders in their firms. Given the high degree of clustering by birth county within
firms, our choice of the designated entrepreneur has little bearing on the analysis in any case. The legal
representative and the largest shareholder belong to the same birth county in over 90% of firms. Even among
the 58% of firms that are established outside the legal representative’s birth county, as many as 74% of the
listed individuals belong to his birth county.17 This is substantially higher than the statistic that would be
14If logωt hits zero, then all agents will want to become entrepreneurs. Entry flows will subsequently plateau. We do not,however, observe such a slowing down in entry flows in the Chinese data in Figure 3.
15Our canonical model is closely related to Munshi’s (2011) model of occupational choice with community networks. Theimportant difference is that the source of forcing variation in his analysis is based on differences in outside options across commu-nities, which is captured by the σ parameter in our model, whereas we exploit exogenous variation in social connectedness acrosscommunities, measured by the θ parameter.
16Citizenship ID’s were first issued in September 1985 and people born after that date are given an ID at birth. Those bornbefore that date were registered in the county or city where they resided at the time. Given the limited opportunities for labormigration in that period and the cost of moving due to the hukou system, almost all rural-born individuals resided in their birth-counties in 1985. The only exceptions were college students, college graduates, and soldiers, but these numbers were small. Thefirst six digits of the citizenship ID thus reveals the county of birth, with few exceptions, even for those born before September1985.
17Among the county-born legal representatives, 42% establish their firm in their birth county, 11% in their birth prefecture butoutside the birth county, 18% in their birth province but outside the birth prefecture, and 29% outside their birth province.
14
obtained by random assignment of listed individuals in the firm’s sector-location, which is just 6%, highlighting
the role played by the birth county in supporting business.18
Table 4 reports tests of the canonical model, corresponding to (5), that estimate the effect of initial entry
on subsequent entry within birth county-sector-locations. Although a single sector-location is available to
all firms in the model, firms from a given birth county will in practice enter multiple sectors and multiple
locations within these sectors. The implicit assumption when measuring entry within sector-locations is that
birth county networks operate independently at that level. Sectors are measured at the two-digit level and
locations are either counties (including the birth county) or urban districts.19 Initial entry is measured in
the 1990-1994 period, when private firms were first starting to emerge in China, and subsequent entry is
measured separately in 2000-2004 and 2005-2009; the five-year window allows for sufficient entry flow at the
birth county-sector-location level. The analysis is restricted to the 1990-2009 period because we will see that
the decline in the ability threshold predicted by the model starts to weaken beyond that point in time. The
benchmark specification in Table 4, Columns 1-2 includes, in addition, birth county-sector fixed effects and
the total number of initial entrants in the sector-location from all origins to capture generalized location-
based agglomeration.20 We see that initial entry from the birth county has a positive and significant effect
on subsequent entry, with this effect growing stronger over time, as predicted by the model. One additional
initial entrant generates seven additional entrants in the 2000-2004 period and nine additional entrants in the
2005-2009 period.
Conditional on the number of initial entrants from the birth county, the total number of initial entrants in
a given sector-location has no effect on subsequent entry from that birth county in that sector-location. This
result provides empirical support for the key assumption in the model that the birth county is the domain
within business networks are organized in China and that these networks operate independently. It also
provides support for the assumption that individual networks cannot influence the price of the product. If the
members of the network could collude (depending on their market share) or there were limits to market size,
then the total number of entrants, conditional on the number of entrants from the birth county, would also be
relevant.21
Although we assume that network spillovers increase the productivity of their members, an alternative
explanation for clustering is that the networks capture economic rents. Bai et al. (2019), for example, describe
how favored firms have superior access to capital allocated by local governments. If local government officials
18The listed individuals must exert effort and contribute in different ways to the firm, and the moral hazard problem thatapplies to the provision of help between firms will also be relevant within the firm. Using the same logic as above, cooperationwithin the firm will be backed by social sanctions in the origin. More effective social sanctions will, in turn, be associated withgreater representation by individuals from the origin in the firm. Appendix Table E.1, Column 1 reports the estimated relationshipbetween the fraction of listed individuals from the legal representative’s birth county and its population density. The coefficienton population density is positive and significant, as expected. Column 2 reports the same relationship for legal representativesborn in cities where, in contrast, the coefficient on population density is negative and significant.
19There are 3,235 counties or urban districts where firms locate in our data. We measure urban population density at the cityrather than the dissaggregated urban district level in our analysis because urban districts were created after 1982. However, firmlocation is always measured at the urban district level.
20All locations which had a positive number of entrants by 2000-2004 and 2005-2009, respectively, for a given birth county-sectorare included in the estimating equation.
21Appendix Table E.2 replicates Table 4 restricting attention to locations outside the birth county. Although the coefficient onthe total number of initial entrants is now statistically significant, it remains positive and an order of magnitude smaller than thecoefficient on the number of entrants from the birth county. If all firms from the same origin collude, but there is competitionacross (origin-based) networks at the destination, then we would expect to see a negative effect of entry from other origins. Whilepricing may be non-competitive in China (see, for example, Brooks et al. (2016)), the origin-based networks do not appear to bedirectly associated with these distortions.
15
Table 4. The Effect of Initial Entry on Subsequent Entry (birth place level)
Dependent variable: subsequent entrants from the birth place
Birth place: county county city
Time period: 2000-2004 2005-2009 2000-2004 2005-2009 2000-2004 2005-2009
(1) (2) (3) (4) (5) (6)
Initial entrants from the birth place 7.120*** 8.935*** 5.198*** 5.723*** 7.385*** 6.668***(0.686) (0.956) (0.972) (1.281) (0.750) (0.799)
All initial entrants at the destination 0.054 -0.020 – – –(0.048) (0.056)
Initial entrants from the birth place ×birth place population density
– – 1.356** 2.277** -0.016 -0.261**
(0.564) (0.937) (0.107) (0.113)Distance to the birth place – – -2.225*** -2.962*** -3.369*** -3.145***
(0.104) (0.106) (0.229) (0.162)
Mean of dependent variable 3.156 3.170 3.156 3.170 4.148 3.579Origin-sector fixed effects Yes Yes Yes Yes Yes YesLocation fixed effects No No Yes Yes Yes YesObservations 384,031 778,897 384,031 778,897 313,915 446,696Note: number of entrants is measured at the birth place- sector-destination level.Initial entry is computed over the 1990-1994 period and sectors are defined at the two-digit level.Number of entrants is obtained from the SAIC registration database and birth place population density is computed from the 1982population census.Standard errors clustered at birth place-sector level are reported in parentheses. * significant at 10%, ** at 5%, *** at 1%.
favor entrepreneurs from their hometown, then this would explain why firms from a given birth county cluster
at the same location.22 However, while pre-existing social ties might confer a temporary advantage, resulting
in additional entry, it will not persist because government officials are frequently rotated precisely to avoid
such corruption. Based on data from city yearbooks and the online CV’s of government officials, we estimate
that the median (mean) term-length for a city mayor in China is 4 (3.7) years. Government favoritism based
on pre-existing social ties, thus cannot explain the long-term effect of initial entry on subsequent entry that
we observe in Table 4.
A related mechanism that might explain the inter-temporal correlation is based on the idea that stronger
networks are more effective at lobbying the government official who is in place. A testable implication of
network competition for scarce government credit, and resources more generally, is that greater entry from
other origins into a particular location will reduce entry from a given origin into that location. We find no
evidence of such negative cross-community spillover effects. While particular entrepreneurs may have favored
access to government credit, the birth county networks do not appear to be facilitating this process. What
they do, instead, is to mobilize capital internally. Recently available data from the Enterprise Survey for
Innovation and Entrepreneurship in China (ESIEC) which uses the registration database as the sampling
frame for a subset of firms indicates that initial registered capital is obtained from the following sources: self
finance (76%), the owners’ social network (15%), and bank loans (9%). More importantly, and in line with
the productivity enhancing role of the network that is assumed in the model, 55% (62%) of the firms report
that their largest stable supplier (buyer) either belongs to their social network or was referred by a member
of the network.23
22Fisman et al. (2018) document the same type of favoritism, based on hometown ties, in the Chinese Academy of Sciences.23The social network is defined to include individuals with pre-existing social ties to the entrepreneur: relatives, laoxiang
16
Although the model assumes that initial entry is exogenously determined, entrepreneurs will select partic-
ular sectors and locations in practice. If initial entrants were more likely to select into persistently favorable
destinations, then a positive correlation between initial entry and subsequent entry could be obtained without
requiring networks to be active. Alternatively, if entrepreneurs tend to locate at proximate destinations, then
a spurious correlation could once more be obtained (because unobserved proximity is fixed over time). We
account for these possibilities in Table 4, Columns 3-4 by adding location fixed effects and the distance between
the birth county and the location under consideration in the estimating equation. Location fixed effects can
be included because entrepreneurs from multiple origins will establish their firms in the same place, but the
variable measuring initial entry from all origins will now be subsumed in the fixed effects. The additional
controls will account for a wide class of omitted variables; what remains is the possibility that some factor
other than distance ties birth counties to specific sector-locations. We address this possibility by testing the
additional prediction of the model, based on the interaction of birth place population density and initial entry.
Table 4, Columns 3-4 reports estimation results with the augmented specification, where we see that the inter-
action coefficient is positive and significant, and increasing from 2000-2004 to 2005-2009, as predicted by the
network model. As a placebo test, we estimate the same equation in Columns 5-6 with city-born entrepreneurs.
The major change in the estimated coefficients, when compared with the results obtained with county-born
entrepreneurs, is that the interaction coefficient is now negative (and significant with 2005-2009 entry as the
dependent variable).
We remain agnostic about the domain of the network in our analysis; i.e. whether it covers the entire birth
county or the clan within the county. Given that clan concentration is increasing in population density in
Chinese counties, the predictions of the model can be tested at the county level even if networks are organized
within clans. Having tested the model at the county level, we now go down to the clan level to directly ascertain
the domain of the network. As with the analysis using CFPS data, we focus on the county-born entrepreneurs
and infer their clan affiliation from their birth county and surname. The predictions of the model are tested at
the clan level by estimating the effect of initial entry in the 1990-1994 period on subsequent entry, separately
in 2000-2004 and 2005-2009, within clan-sector-location. The estimating equation includes birth county-sector
fixed effects as well as the total number of initial entrants from the birth county into that sector-location
(to allow for the possibility that network spillovers extend beyond the clan to others from the same birth
county). The coefficient on the number of initial entrants, at the level of the clan and the county, is positive
and significant in Table 5, Columns 1-2, although the former is an order of magnitude larger. As predicted,
the initial effects grow larger over time; i.e. from 2000-2004 to 2005-2009. Columns 3-4 interact initial entry
with the population density of the birth county. The initial effects are larger for higher population density
birth counties, at the level of the clan and the county, in 2000-2004 and in 2005-2009, and increasing over
time, once again as predicted by the model.
The results in Table 5 indicate that spillovers are concentrated within the clan, although they do extend to
a limited extent beyond its boundaries to other entrepreneurs in the same county from other clans. Assuming
that the spillovers operate exclusively within the clan, our model generates independent predictions for the
(homeplace fellows), classmates, friends, army comrades, colleagues, and past business partners. These categories are not exclusive;in particular, relatives, classmates, and many individuals from the other categories will be laoxiang.
17
Table 5. The Effect of Initial Entry on Subsequent Entry (clan level)
Dependent variable: subsequent entrants from the clan
Time period: 2000-2004 2005-2009 2000-2004 2005-2009
(1) (2) (3) (4)
Initial entrants from the clan 3.275*** 4.082*** 2.422*** 2.689***(0.325) (0.353) (0.304) (0.305)
All initial entrants at the destinationfrom the birth county 0.031*** 0.051*** 0.012 0.006
(0.007) (0.012) (0.013) (0.017)Initial entrants from the clan × birthcounty population density – – 0.650** 1.092***
(0.258) (0.331)Initial entrants from the birth county× birth county population density
– – 0.012* 0.029**
(0.007) (0.013)Distance to the birth county – – -0.174*** -0.259***
(0.012) (0.015)
Mean of dependent variable 1.381 1.397 1.381 1.397Origin-sector fixed effects Yes Yes Yes YesLocation fixed effects No No Yes YesObservations 866,344 1,723,509 866,344 1,723,509
Note: number of entrants is measured at the clan-sector-location level. Clan is defined by the same surname and the same birth county.Initial entry is computed over the 1990-1994 period and sectors are defined at the two-digit level.Number of entrants is obtained from the SAIC registration database and birth place population density is computed from the 1982population census.Standard errors clustered at birth place-sector level are reported in parentheses. * significant at 10%, ** at 5%, *** at 1%.
concentration of business activity within clans drawn from the same birth county, across counties with different
population densities and over time. As discussed above, the total expected sanction and, hence, the level of
cooperation in clan j can be expressed as the product of two functions: ν1(p), which is increasing in population
density p, and ν2(ζj), which is increasing in the relative size of the clan. When entrepreneurs remain connected
to their birth counties, this formulation applies to the incentive compatible level of help that can be supported
in the business network. The θ parameter, which measures this help, can thus be approximated by the product
of two functions when it is measured at the level of the clan: θ1(p) and θ2(ζj).24 Now suppose that there are
two clans in the county: a majority clan, accounting for a fraction M > 12 of the population, and a minority
clan accounting for a fraction m ≡ 1 −M of the population. With two clans, the clan concentration among
entering entrepreneurs (measured by the Herfindahl Hirschman Index) will be locally monotonically increasing
in the entry share of the majority clan or, equivalently, in the ratio of its share to the share of the smaller
clan. From equation (5):
eMt
emt=
[1 + θ1(p)θ2(M(p))
1 + θ1(p)θ2(m(p))
]t−1
(6)
We know from the analysis with CFPS data that clan concentration is increasing in p. Hence, M(p) is
increasing in p, while m(p) is decreasing in p. We can then infer the following from equation (6): (i) Clan
concentration among entering entrepreneurs is increasing over time (because the term in square brackets is
greater than one). (ii) Clan concentration is increasing in population density at each point in time. (iii) The
24The reason why we could express the θ parameter, at the level of the county, as a function of p alone is because clanconcentration is increasing in p and, hence, θ2(ζ) aggregated to the county level is increasing in p.
18
positive relationship between clan concentration and population density will grow stronger over time.
Figure 4. Clan Concentration and Population Density
0
3
6
adju
sted
Her
finda
hl H
irsch
man
Inde
x
0 .02 .04 .06 .08 .1
Birth county population density
1990-1994 1995-1999 2000-2004 2005-2009
Time period:
Source: SAIC Registration database and 1982 population census.Clan concentration measured by the Herfindahl Hirschman Index (HHI) across surnames of entering entrepreneurs from a given birthcounty, divided by the expected HHI that would be obtained by random assignment, given the number of entrants and the number ofsurnames at each point in time.
Table 6. Clan Concentration and Population Density
Dependent variable: adjusted HHI across clans
Time period: 1990-1994 1995-1999 2000-2004 2005-2009
(1) (2) (3) (4)
Birth county population density 0.072*** 0.367*** 0.583*** 0.596***(0.013) (0.029) (0.042) (0.054)
Mean of dependent variable 0.830 1.660 3.291 4.790Observations 1,610 1,624 1,624 1,624
Note: clan is defined by the same surname and the same birth county.Clan concentration measured by the Herfindahl Hirschman Index (HHI) across surnames of entering entrepreneurs from a given birthcounty, divided by the expected HHI that would be obtained by random assignment, given the number of surnames (clans) and the numberof entrants from the birth county at each point in time.Population density is measured in units of 10,000 people per square km, and then converted to Z-score.Control variables include population, education and occupation distribution in the county.Population density, population, education and occupation distribution are computed from the 1982 population census.Robust standard errors are reported in parentheses. * significant at 10%, ** at 5%, *** at 1%.
Figure 4 reports the relationship between the clan concentration of entering entrepreneurs, measured by
the Herfindahl Hirschman Index (HHI), and population density in their birth county at each point in time. The
HHI statistic is adjusted for the fact that measured concentration could vary with the number of entrepreneurs
and the number of clans just by chance, using a normalization derived in Appendix A.25 We will use the same
adjustment for all the analyses of concentration that follow. Clan concentration among entering entrepreneurs
is increasing over time and increasing in birth county population density at each point in time, precisely as
25Previous attempts to examine the spatial distribution of production; e.g. Ellison and Glaeser (1997), Duranton and Overman(2005), have also taken account of this feature of all concentration statistics.
19
our model of within-clan-network-based cooperation would predict. Although it is difficult to visually assess
whether the relationship is growing stronger over time, we see in Table 6, which reports parametric estimates
corresponding to Figure 4 in each time period, that the coefficient on population density is positive and
significant in each time period and increasing over time.26
3 The Augmented Model: with Sector-Location Choice and Capital In-vestment
In this section we extend the canonical model to incorporate sector-location choice and capital investment. One
advantage of the augmented model is that it generates predictions for additional outcomes; the distribution of
firms across sectors and across locations within sectors, as well as firm size. A second advantage is that once
the structural parameters of the augmented model are estimated, it can be used to quantify the impact of the
origin-community based networks on both aggregate firm entry and the stock of capital at a critical stage in
the growth of the Chinese economy. While the predictions of the canonical model were derived with respect to
initial capital, to emphasize the inter-temporal link between firms from the same birth county, we now derive
robust (but less direct) predictions for changes in outcomes over time and, more importantly, for changes
in the relationship between birth county population density and these outcomes over time, when community
networks are active.
There are multiple destinations Bi, i = 1, 2.. for entrepreneurship. The destinations denote sectors or
locations. For simplicity we assume these destinations are ex ante symmetric, except for entry at the initial
date. In destination Bi at date t, an entrepreneur with ability ω selects capital size K, and has a production
function
y = Aitω1−αKα (7)
where α ∈ (0, 1) is the capital elasticity, and Ait denotes CTFP in destination i which takes the form (as
explained in the previous section):27
Ait = A0 exp(θ(p)ni,t−1) (8)
where ni,t−1 denotes incumbent stock from the origin or birth county in destination i at the end of t− 1. As
before, social connectedness θ(p) is increasing in population density in the birth county.
The dependence of CTFP on the size of the incumbent stock represents one source of network complemen-
tarity, reflecting gains from intra-network cooperation in improving productivity for those who have already
entered sector i. We add to this a second source of network complementarity, which pertains to ‘referrals’ or
‘access’ to particular business sectors or locations. A fixed fraction k ∈ (0, 1) of new agents in every cohort
receive an opportunity to become an entrepreneur. Within this group of ‘potential entrants’, the fraction that
get an opportunity to enter destination Bi equals si,t−1, the share of incumbent entrepreneurs from the origin
community already in that destination. This reflects the formation of aspirations, access to information, or
referrals provided by older members from the same origin community in a given destination.
26Pooling data from all time periods, we see in Appendix Table E.3 that the time period coefficient and the population density-time period coefficient are both positive and significant. This result is robust to restricting the sample to firms established outsidetheir entrepreneur’s birth county.
27The A0 term incorporates the product price and labor productivity. Labor is not included as a variable input in the productionfunction because it is not observed in our data. With the Cobb-Douglas specification of the production function, the optimal laborinput can be derived as a function of the model’s parameters and is subsumed in the A0 term.
20
Apart from the decision of whether or not to enter a given destination when presented with the opportunity,
an agent decides on how much capital to invest. All agents incur the same cost of capital r which is exogenous
and fixed across all t and all origins. We are thus abstracting from possible network complementarities
operating via internal capital markets, as in Banerjee and Munshi (2004), which arise in response to financial
market imperfections. To the extent that larger and higher quality networks lower borrowing costs for their
members, the resulting dynamics turn out to be very similar to those generated via productivity spillovers,
and would thus amplify the dynamics generated by the latter alone.28 We also assume a fixed price of the
product, unaffected by supply from the network. This abstracts from price collusion among network members,
as well as limits to market size in a competitive context. These seem plausible in the Chinese setting, where
most sectors are comprised of a large number of origin county networks, and both domestic and international
market opportunities are large.29
3.1 Occupational Choice
To determine occupational choice, we first calculate the profits a new agent in any cohort with a given ability
ω expects to earn upon entering a given business destination (sector and location) when the CTFP in that
destination is expected to be A. The latter is a sufficient statistic for the specific date, destination in question,
and existing network size and quality (which determine CTFP as per (8)). The optimal capital size K must
maximize Aω1−αKα − rK, and thus satisfies:
logK(ω,A) = logω + log φ+1
1− αlogA− 1
1− αlog r (9)
(where φ ≡ α1
1−α ). The resulting profit satisfies
log Π(ω,A) = logω + logψ +1
1− αlogA− α
1− αlog r (10)
(where ψ ≡ φα − φ).30
Of the new agents receiving an offer, the ones that will decide to enter business are those who receive a
higher profit in that destination compared to the traditional occupation.31 These agents will be endowed with
a level of ability that exceeds a threshold ω:
logω > logω ≡ 1
1− σ[log
1
ψ− 1
1− αlogA+
α
1− αlog r] (11)
As with the canonical model, we assume that the threshold lies in the interior of the ability distribution at the
beginning of the process for each destination, and we will restrict attention to ‘early phases of industrialization’
when this continues to be true.28We ignore the role of labor networks in the model. The owner of the firm and the workers rarely belong to the same community,
even in network-based economies. The historical and contemporary experience, across the world, indicates that incumbent workers(with a reputation to maintain within their firms) are the primary source of job referrals.
29Based on the registration data, firms from a given origin county account for 13% of firms at the destinations where they locate,on average (within narrow two-digit sectors). This statistic is based on all entrepreneurs, including those who locate their firmsin their county of birth. This limited market power explains, in part, why we failed to observe negative cross-community entryeffects when testing the canonical model.
30If we allowed for credit networks organized around the origin county and parameterized the interest rate as r =r0exp(−η(p)ni,t−1), then the productivity channel operating through the A term and the credit network channel would notbe separately identified. Although the model is set up so that networks operate through the productivity channel, all the resultsthat follow would go through if, instead, they operated through the credit channel.
31Recall that the profit in the traditional occupation was specified to be ωσ, where σ ∈ (0, 1), in the canonical model. We retainthis specification in the augmented model.
21
Notice that agents receiving an entrepreneurial opportunity make their decision selfishly and myopically.
The former assumption implies that they ignore the consequences of their entry decisions on the profits of
other agents. The latter states that they make their choice solely to maximize their date–t profits, ignoring
consequences at later dates. This enables us to compute the entry dynamics recursively, simplifying the
analysis considerably. If agents were more far-sighted, they would have to forecast current and future levels of
entry from the same origin county, generating strategic complementarity of entry decisions within each cohort.
This extension is considered in Appendix B, where entry decisions at t are based on the discounted sum of
profits at t and t+1, rather than t alone. We show there under some natural conditions that a unique rational
expectations equilibrium exists, whose comparative statics are similar to those in the simpler myopic model.
If anything, the myopic model generates a conservative bias in entry decisions. This is because a network’s
size cannot ever decrease over time and its quality does not change, and neither do profits in the traditional
sector. Those deciding to enter based on a myopic calculation would also want to enter if farsighted, while
some others deciding to stay out on myopic grounds may wish to enter when they anticipate future network
growth, which would further raise the returns to entrepreneurship.
3.2 Dynamics of Entry and Concentration
The different business destinations have identical ‘fundamentals’. At the beginning of the process (t = 0),
there is a small, exogenous number ni0 of older entrepreneurs (from cohorts preceding t = 1) who have already
entered Bi. These represent the initial conditions for the dynamics. These historical entry levels will generically
not be exactly balanced across destinations; without loss of generality suppose ni0 > ni−1,0 > 0 for all i. We
show first that the initial imbalance across destinations will cumulate thereafter, with entrants in later cohorts
increasingly locked-in to the destinations with higher initial presence.
To derive entry in subsequent cohorts, we start with the threshold condition (11), which determines the
measure of agents from cohort t who would choose to enter destination Bi if they had the opportunity.
Combining this with the fraction ksi,t−1 of those agents that have an opportunity to enter, we derive the
volume of entry eit in cohort t into Bi as a function of the state variables ni,t−1, si,t−1:
eit = ksi,t−1[B + Cθ(p)ni,t−1]
where B ≡ 1 − 11−σ log 1
ψ −α
(1−σ)(1−α) log r + 1(1−σ)(1−α) logA0 and C ≡ 1
(1−σ)(1−α) . This expression reduces
further to
eit = Lsi,t−1 + κ(p)Nt−1s2i,t−1 (12)
where L denotes kB; κ(p) denotes Ckθ(p) which is rising in p, and Nt−1 ≡∑
i ni,t−1 denotes the aggregate
number of business entrepreneurs from past cohorts from the same origin. Aggregating (12) across sectors, we
obtain an expression for the dynamics of aggregate entry:
Nt −Nt−1 ≡ Et ≡∑i
eit = L+ κ(p)Nt−1Ht−1 (13)
where Ht−1 ≡∑
i s2i,t−1 denotes the Herfindahl Hirschman Index for concentration at t− 1. Equations (12,13)
define the dynamics of the vector (Nt, sit, i = 1, 2..), where sit ≡ si,t−1Nt−1
Nt+ eit
Nt.
22
Proposition 1 Concentration Ht and aggregate entry flow Et are both rising in t.
The proofs of this proposition, as well as subsequent propositions, are provided in Appendix C. The intuitive
reason for concentration to rise over time is simple: a destination with higher incumbent stock is both more
profitable and generates more opportunities for entry, so its share grows faster, reinforcing the higher initial
presence. Note that the compounding effect of initial entry, due to network spillovers, which resulted in an
increase in entry over time in the canonical model, is present here as well. The network complementarity
associated with post-entry productivity spillovers, embodied in the Nt−1 term in (13), is now reinforced by
the additional network complementarity associated with the referrals; i.e. the Ht−1 term. Entry Et will rise
over time from (13) if concentration is increasing over time.
The compounding network effect is stronger for firms from higher p origins, on account of the κ(p) multiplier,
so one would also expect the level of concentration and entry, and their growth over time, to be rising in p.
Verifying this conjecture is more complicated, however, especially with respect to concentration. To illustrate
this, consider the case of two destinations i = 1, 2, whereupon concentration is monotone increasing in the
share of the destination with a higher initial presence. We can then focus on the dynamics of the share of this
initially dominant destination, which (without loss of generality) we denote by destination 1:
s1t ≡ [Nt
n1t]−1 = [
L+Nt−1 + κ(p)Nt−1Ht−1
Ls1,t−1 + n1,t−1 + κ(p)Nt−1s21,t−1
]−1
= [1 + (1
s1,t−1− 1){L+Nt−1 + κ(p)Nt−1(1− s1,t−1)
L+Nt−1 + κ(p)Nt−1s1,t−1}]−1 (14)
Proposition 2 With two destinations:
(a) Entry Et and concentration Ht are rising in p (at any given t).
(b) Et − Et−1 and Ht −Ht−1 are both rising in p, if κ(p) < 1 for all p and the share of the larger sector at
t− 1 is not too close to 1 (e.g., below 34).
Part (a) confirms that the level of concentration is rising in p at any t, which in turn implies the same
for entry flows. Part (b) shows that growth of entry or concentration is rising in p at ‘early stages’ of the
industrialization process; i.e. when concentration is not too high. The qualifier is required because the share
of the dominant sector B1 is bounded above by 1. Thus, the share of the dominant destination cannot be
increasing faster forever; eventually, its growth rate will flatten out as the share approaches one.
The results concerning the dynamics of concentration across destinations translate into testable predictions
concerning either sectoral or spatial concentration, given that destinations correspond to sectors or locations.
We partition the set of destinations into sectors, with each sector consisting of a subset of locations. Proposition
1 can then be extended to show that spatial (location) concentration within any given sector must be rising in
t. The same can be shown for sectoral concentration, provided that sectors with higher initial shares are also
characterized by higher intra-sectoral spatial concentration at date 0.32 Moreover, with two locations within
any sector, the results on concentration in Proposition 2 apply across sectors or across locations within sectors.
32The reason is that the expression for entry flow into sector c is modified to ect = κLsc,t−1 + κNt−1s2c,t−1Hc,t−1, so the term
involving the quadratic term in lagged sectoral share is weighted by lagged intra-sectoral spatial concentration Hc,t−1.
23
3.3 Ability Selection and Firm Size Dynamics
Next we derive predictions concerning entrepreneurial ability and firm size. We first show that network effects
generate negative selection on ability: from (11), as CTFP increases over time, the threshold for entry falls,
and entrepreneurs with lower ability start entering. This negative selection has consequences for firm size.
Substituting from (11) in (9), initial capital of the marginal entrant is decreasing in CTFP, Ait:
logKmit = U ′ − σ
(1− σ)(1− α)logAit (15)
where U ′ ≡ log φ − 11−σ logψ − 1
1−α log r, and logAit = logA0 + θ(p)Nt−1si,t−1. The negative selection on
individual ability that accompanies a stronger network (with higher CTFP) outweighs its productivity benefit.
The marginal entrepreneurs that enter are thus less productive and enter with smaller firm sizes.33 The same
argument applies to comparisons at any given t across different p origins: marginal entrants from higher p
origins, with stronger networks, enter with smaller firm sizes. If σ ∈ (12 , 1) this is true also for the average
entrant: firms from high p origins enter with smaller initial capital on average, with the opposite result
holding if σ < 12 .34 To see this, observe that substituting from (11) in (9), the capital size of the average
entrant satisfies:
logKait = W +
1− 2σ
2(1− α)(1− σ)logAit (16)
where W ≡ log φ + 12 + 1
2(1−σ) log 1ψ −
2−α−2σ2(1−α)(1−σ) log r. All firms face the same cost of capital and there are
no mark-ups in our model. The preponderance of small and seemingly unproductive firms often noted in
developing countries, which is typically attributed to wedges in factor prices and mark-ups in output price in
the misallocation literature, may instead just be a manifestation of strong network effects! Our model implies
that their own productivity understates their contribution via spillovers to their network.
In contrast to the results for initial capital, post-entry growth rates of firm size for any given cohort can
be shown to be rising in p and over time. (9) implies that the capital at date t′ > t of a cohort t entrepreneur
with ability ω is given by
logKaitt′ = logω + log φ− 1
1− αlog r +
1
1− α[logAit + θ(p)
t′−1∑l=t
eil] (17)
implying a growth rate at period t′:
logKait,t′ − logKa
it,t′−1 =1
1− αθ(p)eit′ (18)
In our model, growth in incumbent firm size is independent of the entrepreneur’s ability and cohort and is
driven entirely by contemporaneous changes in CTFP. Once we average across destinations, these changes in
CTFP are measured by aggregate entry flows of firms from the origin.
Proposition 3 With two destinations:
33This result does not depend on assumptions concerning the distribution of ability. To see this, observe that expressions (9, 10)show that capital size and entrepreneurial profit depend on individual ability and CTFP in exactly the same way. The marginalentrepreneur must be of lower ability when CTFP is higher, and must be indifferent between the traditional occupation andentrepreneurship. Profits will thus be lower in the traditional occupation for an agent with lower ability. So the same is true forentrepreneurial profit, and hence for capital size.
34This depends on the assumption of a log uniform distribution of ability.
24
(a) Averaging across destinations, ability and initial capital of marginal entrants (also of average entrants if
σ > 12) is decreasing in t (for any given p) and in p (for any given t), and decreasing more steeply in p
across successive cohorts.
(b) Averaging across destinations, the growth rate of capital of incumbent entrepreneurs of any past cohort t
from t′ − 1(> t) to t′ is rising in t′ and in p (more steeply over time).
From (11) and (15), the marginal entrant’s ability and initial capital are decreasing in logAit. From (16), the
average entrant’s ability and initial capital are also decreasing in logAit if σ > 12 . logAit ≡ logA0 + θ(p)ni,t−1
is increasing in Nt−1 when it is averaged across destinations. From Propositions 1 and 2 we know that Et and,
hence, Nt is increasing in t (for any p), increasing in p (for any t), and increasing more steeply in p over time.
Hence, part (a) of Proposition 3 follows immediately. A similar argument can be used to establish part (b).
Averaging across destinations, eit′ is replaced by Et′ on the right hand side of equation (18). The result then
follows from Propositions 1 and 2. Firms from high-p origins start smaller, but subsequently grow faster.35
This dual prediction will be especially helpful in distinguishing our network-based model from alternative
models, as discussed in the next section.
3.4 Alternative Explanations
To what extent do the preceding results rely on network spillovers? Could they be obtained, instead, by
relaxing different assumptions of our model, while shutting down the network component? These questions
are relevant because although population density is plausibly associated with social connectedness in the birth
county and, hence, network quality, it could also be correlated with other factors that independently determine
the dynamics of entry, concentration, and firm size. The discussion that follows systematically examines this
possibility by introducing new sources of (possibly time-varying) heterogeneity at the origin, which are, in
turn, correlated with population density, and by allowing firms from different origins to have favorable access
to destinations of varying quality. Our model treats sectors and locations interchangeably. Because locational
heterogeneity is an important alternative that we must consider, entrepreneurs choose between locations (which
we refer to as destinations for expositional convenience) rather than sectors in the models that follow.
3.4.1 Origin Heterogeneity
Our model assumes that the stock of potential entrepreneurs, k, is constant across origin counties and cohorts.
Suppose that we relax this assumption and let k(p, t) be a twice differentiable function satisfying kp > 0, kt >
0, kpt > 0. This could be because higher population density counties simply have larger populations that are
growing relatively fast over time or because their residents have greater wealth or preferred access to finance,
which facilitate entry into business. An additional source of origin heterogeneity could be in payoffs in the
traditional occupation across counties. Our model assumes that the payoff, ωσ, where ω is individual ability,
35In a related paper, Banerjee and Munshi (2004) find that outsiders in Tirupur’s garment cluster, who face a higher cost ofcapital because they have weaker local credit networks, start with smaller firms and then grow faster (because they are positivelyselected on ability). To explain Banerjee and Munshi’s findings, our model would need to be augmented to allow firm growth to beincreasing in the entrepreneur’s ability, with the additional condition that the ability effect needs to dominate the network effect(which is stronger for the insiders).
25
is the same in all counties and constant over time. However, the payoff could be lower in higher population
density counties because there is a larger population for a given amount of resources (such as agricultural
land). It is also possible that this population pressure is increasing over time. We allow for this possibility
by representing the payoff in the traditional sector by ωσv(p, t), where v(p, t) is a twice differentiable function
satisfying vp < 0, vt < 0, vpt < 0.36
Abstract for the time being from heterogeneity across destinations, so the TFP of any entrepreneur with
ability ω at any destination at time t is ω1−αAt, with At growing exogenously over time. Moreover, an
exogenous share si of potential entrepreneurs at each origin have the opportunity to enter any given destination.
Owing to the absence of network effects, neither At nor si depend on p.
The ability threshold for entry into destination i from an origin with population density p in this alternative
model would equal
logωi(p; t) =1
1− σ
[log
1
ψ+
α
1− αlog r − 1
1− αlogAt + log v(p, t)
](19)
while the expression for entry flows is:
ei(p, t) = sik(p, t)
[1 + Z +
1
(1− α)(1− σ)logAt −
1
1− σlog v(p, t)
](20)
where Z ≡ 11−σ logψ+ α
(1−α)(1−σ) log r. The entry flows will be rising in p and in t, and the slope with respect
to p will be rising in t. The alternative model can thus generate our model’s predictions for entry. However, the
share of different destinations will be constant and independent of p. In order to obtain the same predictions
for spatial concentration generated by the network model, the shares si of different destinations would have
to (exogenously) depend on p and t in a way that exactly delivers these results. Although we do not explicitly
incorporate sectors in the alternative model, it would similarly need to be augmented to exactly match our
model’s predictions for the dynamics of sectoral concentration.
With σ ∈ (12 , 1) the initial capital of entrants would fall over time due to the increase in At and the decline
in v(p, t), and would also be falling in p (more steeply over time) due to the v(p, t) term. However, post-entry
growth of firm size would be driven entirely by rising productivity at the destinations, At, which does not vary
with the origins of entrepreneurs. Hence this model would not generate result (b) of Proposition 3 concerning
post-entry growth in firm size across origin counties.
3.4.2 Destination Heterogeneity
Now consider the implications of varying productivity levels and growth rates across destinations. This could
reflect the effect of geography, support provided by local governments (through credit and infrastructure), or
agglomeration spillovers. The latter depend on the total number of firms at a destination, regardless of their
origin. Let Ait denote productivity at destination i at t, which does not vary with the origins of entrepreneurs
in the absence of network effects. Suppose in addition that high p origins have better, and increasing, access
to the faster growing destinations. For instance, if there are two destinations and productivity at destination
36An alternative interpretation of v(p, t) is that it represents the payoff from origin-based networks operating in the traditionalsector.
26
1 is higher and growing faster than at destination 2, then the share s1(p, t) is increasing in t and in p (more
steeply over time). The expressions for entry thresholds and entry flows are now
logωi(p; t) =1
1− σ
[log
1
ψ+
α
1− αlog r − 1
1− αlogAit + log v(p, t)
](21)
ei(p, t) = si(p, t)k(p, t)
[1 + Z +
1
(1− α)(1− σ)logAit −
1
1− σlog v(p, t)
](22)
This model which incorporates both origin and destination heterogeneity would generate the same predic-
tions as Proposition 2 for entry and concentration. There would be greater total entry from high p origins
owing to the origin heterogeneity, coupled with greater access to the faster growing destination. Concentration
would rise over time for entrepreneurs from all origins, owing to faster entry growth into destination 1. This
would be more pronounced for the high p origins, so concentration would rise in p and p ∗ t. Entry thresholds
from high-p origins would be lower due to higher Ait (averaged across destinations) or lower v(p, t), so the
initial capital size result in part (a) of Proposition 3 would also go through. The average rate of growth of
firm size (where we average across destinations) would be higher for high-p origins, owing to their preferred
access to the faster growing destination.
The alternative model specified above can generate the predictions of our model relating to the dynamics
of entry, concentration, and firm size because the key si(p, t), Ait terms are exogenously specified to match
the endogenous evolution of these terms in our model. If firms from each origin locate at a unique set of
destinations, then our network-based model would not be distinguishable from the alternative model with
destination heterogeneity. In practice, however, firms from multiple origins will locate at the same destination.
Destination-time period dummies can then be included in the estimating equation. Conditional on these
dummies, the network model would imply that firms from higher-p origins will grow faster on average because
their growth is driven by changes in CTFP. In contrast, there is no relationship between firm growth and p
in the alternative model once destination-time period dummies are included because there is no longer any
variation within destinations.
One way to incorporate heterogeneity within destinations, without networks, would be to allow firm growth
to vary with the entrepreneur’s ability; this is not a feature of our model. A positive relationship between p in
the origin county and firm growth would then be obtained even within destination-time periods if entrepreneurs
from higher p origins have higher ability on average. However, this model would not explain why firms from
higher p origins, with higher ability, nevertheless have lower initial capital. An alternative model that may be
considered, is that entrepreneurs do not have access to external credit and have to be entirely self-financing
(Song et al., 2011). Suppose that for some reason entrepreneurs from high p counties have a higher shadow
cost of capital, so entering firms start with lower capital size, and thereafter grow faster owing to convergence
forces akin to those in the Ramsey-Solow neoclassical growth model. This model would not be able to explain
the positive relationship between population density and either entry or concentration; high p origins ought
then to be associated with smaller entry flows. Nor would it be able to explain why the positive relationship
between firm size growth and population density is robust to controlling for initial capital size (as shown
below).
27
4 Testing the Augmented Model
4.1 Evidence on Firm Entry
The model predicts that firm entry is (i) increasing in origin county population density at each point in time,
(ii) increasing over time, and (iii) increasing more steeply in population density over time. This is a statement
about the flow of firms rather than the stock. As with the tests of the canonical model, we will thus measure
entry in five-year windows from 1990 to 2009 to ensure that there is a sufficient flow of firms in each time
period. Figure 5 reports nonparametric estimates of the relationship between the entry of firms from each
birth county in each time period and 1982 population density.37 The entry patterns in the figure are visually
consistent with the model’s predictions.38
Figure 5. Firm Entry and Population Density
0
2
4
6
8
Num
ber
of e
nter
ing
firm
s (t
hous
ands
)
0 .02 .04 .06 .08 .1
Birth county population density
1990-1994 1995-1999 2000-2004 2005-2009
Time period:
Source: SAIC registration database and 1982 population census.
Table 7, Columns 1-4 report parametric estimates corresponding to Figure 5, separately by time period.
This allows us to statistically validate the prediction that entry is increasing in birth county population density
at each point in time. As noted, all analyses that estimate the direct effect of birth county population density
will include population, education, and occupational structure (also measured at the county level in 1982)
in the estimating equation. This allows for the possibility that population density is correlated with county
characteristics that independently determine the outcomes of interest. We see in Table 7, Columns 1-4 that
the population density coefficient is positive and significant in each time period. Notice also that the mean of
the dependent variable and the population density coefficient are increasing across time periods, in line with
37The advantage of using a predetermined measure of social connectedness in all of the analysis is that it avoids the possibilitythat changes in population density in later time periods are generated by endogenously determined network-based migration.Nevertheless, as documented in Appendix Figure D.2 using successive rounds of the population census, the ranking of countieswith respect to population density is invariant over time.
38Appendix Figure D.3a reports the corresponding nonparametric relationship between population density in the birth countyand the stock of firms (measured at the end of each time period). The predictions of the model apply to both firm entry; i.e. theflow and the stock of firms. In practice, however, the stock will also take account of exits, which play no role in the model. We seein Figure D.3a that the model’s predictions for the stock of firms go through as well, despite the exits. As an additional robustnesstest, Appendix Figure D.3b reports the nonparametric relationship between population density in the birth county and firm entry,restricting attention to firms that locate outside the birth county. Although the entry result is based on all locations, we see thatthe predictions of the model hold up with this reduced sample of locations as well.
28
predictions (ii) and (iii) above. Formal tests of these predictions are reported later in this section.
Table 7. Firm Entry and Population Density
Dependent variable: number of entering firms number of entering firms
Time period: 1990-1994 1995-1999 2000-2004 2005-2009 1990-1994 1995-1999 2000-2004 2005-2009
(1) (2) (3) (4) (5) (6) (7) (8)
Birth county populationdensity
0.013*** 0.092*** 0.289*** 0.448*** 0.014** 0.118*** 0.382*** 0.575***
(0.003) (0.013) (0.035) (0.052) (0.006) (0.037) (0.101) (0.126)
Mean of dependent variable 0.0306 0.208 0.787 1.560 0.0725 0.483 1.673 3.024Sector fixed effects No No No No Yes Yes Yes YesLocation fixed effects No No No No Yes Yes Yes YesObservations 1,624 1,624 1,624 1,624 1,085,169 1,085,169 1,085,169 1,085,169
Note: number of entering firms from each birth county in each time period is measured in thousands.Population density is measured in units of 10,000 people per square km, and then converted to Z-score.In Columns 5-8, for a given birth county, all sectors and locations that ever have entrants are included in all time periods (assigned zeroentry where necessary). To adjust for differences in the number of sectors and locations across birth counties, the number of entrants ismultiplied by the number of sectors × the number of locations.Control variables include population, education and occupation distribution in the birth county.Population is measured in millions and education is measured by the percent of the population that is literate.Occupation distribution is measured as the share of workers in the birth county in agriculture and industry. Service is the excludedcategory.Number of firms is obtained from the SAIC registration database and population density, population, education and the occupationdistribution are derived from the 1982 population census.Robust standard errors are reported in parentheses. * significant at 10%, ** at 5%, *** at 1%.
As discussed above, both origin heterogeneity and destination heterogeneity can explain the entry results
without requiring origin-based networks to be active. We account for important elements of origin heterogeneity
by including additional birth county characteristics in the estimating equation. We next account for the
possibility that entrepreneurs born in higher population density birth counties have access to faster growing
destinations. Given that firms from multiple origin counties will enter each destination, we can flexibly
accommodate this possibility by including destination fixed effects in the estimating equation. The estimating
equation in Table 7, Columns 5-8 includes sector fixed effects and location fixed effects together with the
birth county characteristics. This equation is estimated separately in each time period and so the fixed effects
capture the changing fortunes of sectors and locations over time.39 Birth county population density continues
to have a positive and significant effect on entry in each time period in Table 7, Columns 5-8. A comparison
of the results obtained with the benchmark specification in Columns 1-4 and the augmented specification in
Columns 5-8 indicates that the inclusion of the fixed effects actually increases the point estimates. This tells
us that entrepreneurs born in high population density counties are selecting sectors and locations that are less
advantageous (receive fewer entrants overall).
4.2 Evidence on Concentration
The model predicts that the concentration of the stock of firms, measured by the Herfindahl Hirschman
Index (HHI) across destinations, is (i) increasing in birth county population density at each point in time,
(ii) increasing over time, and (iii) increasing more steeply in population density over time. Destinations are
defined by sectors or by locations within sectors. Figure 6a reports nonparametric estimates of the relationship
39Entry in Table 7, Columns 5-8 is measured at the birth county-sector-location level in each time period. The number ofentrants is thus multiplied by the county-specific product of the number of sectors and the number of locations so that thedependent variable reflects entry at the level of the county.
29
between sectoral concentration at the two-digit level and 1982 population density in the birth county in five-
year intervals from 1994 to 2009. The HHI is based on the stock of existing firms (net of exits) and, as with
the analysis of clan concentration, is adjusted for the fact that measured concentration could vary with the
number of firms and the number of sectors just by chance, using a normalization derived in Appendix A. The
adjusted HHI is evidently increasing in population density at each point in time and increasing over time,
although it is difficult to visually assess whether the slope of the relationship gets steeper over time. Figure
6b reports nonparametric estimates of the relationship between spatial concentration, within one-digit sectors,
and birth county population density in five-year intervals. Although the model assumes that all destinations
are symmetric, one obvious asymmetry in practice is that transportation costs are lower when the entrepreneur
chooses to stay back home. We avoid this asymmetry by including all locations in the analysis, measured at
the county or urban district level, except for the birth county. As with the analysis of sectoral concentration,
the spatial concentration within each sector for a given birth county is based on the stock of firms (net of
exits) and is adjusted for the number of firms and the number of external destinations, which would generate
variation in the measured HHI just by chance.40 Matching the predictions of the model, the spatial HHI is
evidently (i) increasing in birth county population density in each time period, (ii) increasing over time, and
(iii) increasing more steeply over time.
Figure 6. Concentration and Population Density
0
3
6
9
adju
sted
Her
finda
hl H
irsch
man
Inde
x
0 .02 .04 .06 .08 .1
Birth county population density
1994 1999 2004 2009
Year:
(a) Sectoral Concentration
1
2
3
adju
sted
Her
finda
hl H
irsch
man
Inde
x
0 .02 .04 .06 .08 .1
Birth county population density
1994 1999 2004 2009
Year:
(b) Spatial Concentration
Source: SAIC registration database and 1982 population census.Sectoral concentration is measured by the Herfindahl Hirschman Index (HHI) across two-digit sectors, divided by the expected HHI thatwould be obtained by random assignment, given the stock of firms and the number of sectors at each point in time.Spatial concentration, within one-digit sectors, is measured by the Herfindahl Hirschman Index (HHI) across destination locations (outsidethe birth county) divided by the expected HHI that would be obtained by random assignment, given the stock of firms and number ofdestination locations at each point in time.
Table 8 reports parametric estimates corresponding to Figure 6a and Figure 6b. The usual birth county
characteristics are included in the estimating equation. Sector fixed effects are also included in the estimating
equation with spatial concentration (within sectors) as the dependent variable to allow for the possibility that
concentration varies independently across sectors (possibly due to the nature of the production technology and
40We measure spatial concentration within one-digit rather than two-digit sectors to allow for a sufficient flow of firms acrosslocations. To maintain consistency across time periods, we only include birth county-sectors that have multiple entrants in all timeperiods. This is not a constraint in the sectoral analysis because all birth counties have multiple entrants in each time period.
30
the associated gains from agglomeration). Population density in the birth county has a positive and significant
effect on (adjusted) sectoral and spatial concentration at each point in time. The mean of the dependent
variable and the population density coefficient are increasing over time, in line with predictions (ii) and (iii),
which we test formally below.
Table 8. Sectoral and Spatial Concentration and Population Density
Dependent variable: adjusted HHI across sectors adjusted HHI across locations
Year: 1994 1999 2004 2009 1994 1999 2004 2009
(1) (2) (3) (4) (5) (6) (7) (8)
Birth county populationdensity
0.106*** 0.417*** 0.444*** 0.574*** 0.014* 0.041*** 0.047* 0.072*
(0.023) (0.057) (0.061) (0.074) (0.008) (0.014) (0.027) (0.040)
Mean of dependent variable 1.039 2.839 4.622 6.065 0.936 1.010 1.295 1.777Observations 1,622 1,624 1,624 1,624 5,450 15,076 23,727 26,769Note: sectoral concentration measured across two-digit sectors and spatial concentration, within one-digit sectors, is measured acrossdestination locations (outside the birth county). Concentration statistics are adjusted for expected concentration due to random assignment.Population density is measured in units of 10,000 people per square km, and then converted to Z-score.Sector fixed effects included in the regression with spatial HHI as the dependent variable.Control variables include population, education and occupation distribution in the birth county.Number of firms is obtained from the SAIC registration database and population density, population, education and the occupationdistribution are derived from the 1982 population census.Robust standard errors are reported in parentheses. * significant at 10%, ** at 5%, *** at 1%.
Table 9 formally tests the model’s predictions for changes in firm entry, sectoral concentration, and spatial
concentration over time and across birth county population density over time. Data from all time periods
are pooled and the estimating equation now includes birth county population density, time period, and the
interaction of these variables. Since the cross-sectional relationship with population density in each time period
has been previously reported with each outcome, we only report the coefficient on the time period variable
and the interaction coefficient. Restricting the sample to county-born entrepreneurs in Table 9, Columns
1-3, the time period coefficient and the interaction coefficient are positive and significant with the number
of entrants, sectoral concentration, and spatial concentration as the dependent variables, as predicted by the
model. As a placebo test, we restrict the sample to entrepreneurs born in cities in Table 9, Columns 4-6.
Population density is not positively associated with social connectedness in cities and thus we do not expect
to find support for the model’s predictions with this set of entrepreneurs. The time period coefficient and the
interaction coefficient are both positive and significant with entry as the dependent variable but, as discussed,
many alternative models can generate this result without a role for birth county networks. The model’s
predictions for concentration are less easy to explain away. Reassuringly, the interaction coefficient for the
city-born entrepreneurs is negative and significant with sectoral concentration as the dependent variable and
statistically indistinguishable from zero (at conventional levels) with spatial concentration as the dependent
variable, contrary to the predictions of our model.
4.3 Evidence on Firm Size
The model predicts that the ability and the initial capital of the marginal entrant is (i) decreasing in birth
county population density at each point in time, (ii) decreasing over time, and (iii) decreasing more steeply in
31
Table 9. Entry, Concentration, and Population Density (time and interaction effects)
Birth place: county city
Dependent variable: number ofentrants sectoral HHI spatial HHI number of
entrants sectoral HHI spatial HHI
(1) (2) (3) (4) (5) (6)
Time period 0.517*** 1.686*** 0.506*** 0.661*** 0.299*** 2.054***(0.016) (0.020) (0.021) (0.026) (0.008) (0.069)
Birth place populationdensity × time period 0.353*** 0.165*** 0.134*** 0.353*** -0.026*** 0.030
(0.029) (0.019) (0.012) (0.041) (0.004) (0.026)
Observations 6,496 6,494 71,022 3,284 3,283 21,046Note: the estimating equation includes, in addition, birthplace population density and a constant term.Number of entering firms from each birth place in each time period is measured in thousands.Sectoral concentration measured by Herfindahl Hirschman Index (HHI) across two-digit sectors divided by the expected HHI that wouldbe obtained by random assignment.Spatial concentration, within one-digit sectors, is measured by the Herfindahl Hirschman Index (HHI) across destination locations (outsidethe birth county) divided by the expected HHI that would be obtained by random assignment. Sector fixed effects are included in thisspecification.Population density is measured in units of 10,000 people per square km, and then converted to Z-score.Time period is an ordinal variable taking value from 1 to 4 corresponding to successive five-year time windows over the 1990-2009 period.Number of entrants and concentration statistics are derived from the SAIC registration database and population density is derived fromthe 1982 population census.Standard errors clustered at birthplace level are reported in parentheses. * significant at 10%, ** at 5%, *** at 1%.
population density over time. If the negative selection on ability that accompanies a stronger network domi-
nates the positive productivity effect of that network for inframarginal firms, then the preceding predictions
apply to average initial capital as well. However, only the positive network productivity effects are relevant
for post-entry growth rates of firm size.
To test the model, we measure marginal ability and initial capital at the level of the birth county-sector
or birth county-sector-location in each time period and then estimate the (average) effect of birth county
population density and its interaction with time on these variables. We begin in Figure 7a by nonparametrically
estimating the relationship between a measure of ability, based on education, of the marginal entrepreneur
in each birth county-sector-time period and population density in the birth county. It is standard practice
to proxy for ability with education, and recent evidence indicates that education is also a good measure of
entrepreneurial ability (Levine and Rubinstein, 2017). In a developing economy, the level of education will vary
across birth cohorts and in the cross-section (across birth counties) for the same level of ability, depending on
the supply of schooling. Our measure of ability is thus the entrepreneur’s percentile rank in his birth county-
birth cohort education distribution.41 The marginal entrant is the entrepreneur who is placed at the bottom
one percentile of the ability distribution among entering entrepreneurs in each birth county-sector-time period.
We see in Figure 7a that the marginal entrant’s measured ability declines over time; from around the 70th
percentile of his birth county-birth cohort education distribution in the 1990-1994 period to just around the
40th percentile in the 2005-2009 period. The relationship between the marginal entrant’s ability and population
41The education distribution is constructed in each county for birth cohorts from 1920 to 1989 in five-year intervals, based ondata from the 2000 population census. Each entrepreneur is assigned to a birth cohort interval based on his birth year, which isavailable from the registration database, and his position in the relevant education distribution is determined on the basis of hiseducation, which is also obtained from the registration database. The coverage for the education variable is not complete in theSAIC registration database, with a significant minority of entrepreneurs not reporting this information. This has no bearing onthe complementary analysis of firm size, which includes all registered firms.
32
density is also negative in each time period and grows steeper over time.42 Notice, however, that there is a
bottoming out by the last, 2005-2009, period. Our model is only designed to capture firm dynamics up to this
point, which is why the empirical analysis does not extend beyond 2009. For the dynamic analysis of negative
selection that follows, and for the structural estimation, the analysis period will be restricted even further to
the 1990-2004 period.
Figure 7. Marginal Ability, Marginal Initial Capital and Population Density
30
40
50
60
70
80
90
Mar
gina
l abi
lity
(adj
uste
d ed
ucat
ion)
0 .02 .04 .06 .08 .1
Birth county population density
1990-1994 1995-1999 2000-2004 2005-2009
Time period:
(a) Marginal Ability
-2.5
-2
-1.5
-1
-.5
0
Mar
gina
l ini
tial c
apita
l (Lo
g)
0 .02 .04 .06 .08 .1
Birth county population density
1990-1994 1995-1999 2000-2004 2005-2009
Time period:
(b) Marginal Initial Capital
Source: SAIC registration database and 1982 population census.The entrepreneur’s ability is measured by his percentile rank in his birth county- birth cohort education distribution (obtained from the2000 population census). The marginal entrepreneur is defined as the individual at the bottom one percentile of the ability distributionamong entering entrepreneurs in a given birth county-sector-time period.Marginal initial capital is defined as the bottom one percentile of the initial capital distribution at the birth county-sector-time period level.
Figure 7b reports complementary nonparametric estimates of the relationship between marginal initial
capital, measured in logs, and 1982 population density in the birth county in five-year windows over the 1990-
2009 period. Marginal initial capital is defined as the bottom one percentile of the initial capital distribution
at the birth county-sector-time period level.43 As predicted by the model, marginal initial capital is decreasing
over time and decreasing in birth county population density in each time period.44
Notice from Figure 7b that the decline in initial capital with birth county population density does not grow
steeper over time (as implied by the model). One reason why this might be the case is because marginal initial
capital within birth county-sector-time periods is effectively averaged across sectors in the figure. Although this
is not a feature of our model, the capital requirement will vary across sectors, and this must be accounted for in
the empirical analysis. Table 10 allows for this by studying the change in the ability of entering entrepreneurs
and their capital investments over time, within birth county-sectors. The analysis is restricted to the 1990-
2004 period because our measure of marginal ability and initial capital both bottom out (and flatten out)
in the 2005-2009 period in Figures 7a and 7b. We see in Table 10, Column 1, which includes birth county-
sector fixed effects, that the marginal entrant is drawn from lower down in his birth county-cohort education
42Appendix Table E.4 reports parametric estimates corresponding to Figure 7a, separately in each time period. These esti-mates indicate that birth county population density has a negative and significant effect on marginal education among enteringentrepreneurs at each point in time.
43The initial capital for a firm is determined by its initial registered capital, which can be recovered from the SAIC registrationdatabase.
44Appendix Table E.4 reports parametric estimates corresponding to Figure 7b, separately by time period. The populationdensity coefficient is negative and significant in each time period.
33
distribution over time and that this decline in our measure of ability is significantly steeper for entrants from
higher population density counties, as predicted by the model. Table 10, Columns 2-3 use the distribution
of initial capital (in logs) in each entering cohort of firms, in five-year windows over the 1990-2004 period,
to identify the marginal entrant (the bottom one percentile) and the average entrant by birth county-sector.
Including birth county-sector fixed effects in the estimating equation, we see that both the marginal entrant’s
initial capital and the average entrant’s initial capital are decreasing significantly over time. Although the
coefficient on the time period-birth county population density interaction is also negative and significant with
the marginal entrant’s initial capital as the dependent variable, the interaction coefficient is positive (albeit
small in magnitude and statistically insignificant) with average initial capital as the dependent variable.
The analysis of firm size thus far has not accounted for location choices, and the possibility that variation
in these choices across birth counties could be driving the results. Table 10, Columns 4-5 thus includes location
fixed effects, in addition to birth county-sector fixed effects in the estimating equation. Initial capital is now
measured at the birth county-sector-location level in each time period.45 Both marginal initial capital and
average initial capital are declining significantly over time, as in Columns 2-3. Moreover, the coefficient on the
time period-birth county population density interaction is now negative and significant with both dependent
variables, as predicted by the model. As with the analysis of firm entry, accounting for location effects only
strengthens our results, indicating that entrepreneurs from higher population density birth counties selected
less advantageous locations where firms had lower access to capital (within sector) on average.
Table 10. Evidence on Negative Selection
Dependent variable: marginalability
marginalinitial capital
averageinitial capital
marginalinitial capital
averageinitial capital
(1) (2) (3) (4) (5)
Time period -17.908*** -0.868*** -0.116*** -0.609*** -0.095***(0.496) (0.012) (0.009) (0.010) (0.008)
Birth county populationdensity × Time period -0.926*** -0.026** 0.002 -0.061*** -0.020***
(0.351) (0.011) (0.007) (0.009) (0.006)
Mean of dependent variable 49.36 -1.744 -0.401 -1.223 -0.374Origin-sector fixed effects Yes Yes Yes Yes YesLocation fixed effects No No No Yes YesObservations 21,028 43,578 43,578 46,417 46,417Note: The entrepreneur’s ability is measured by his percentile rank in his birth county- birth cohort education distribution (obtained fromthe 2000 population census). The marginal entrepreneur is defined as the individual at the bottom one percentile of the ability distributionamong entering entrepreneurs in a given birth county-sector-time period.Initial capital (in million Yuan) is measured in logs. Marginal initial capital defined by the bottom one percentile of the initial capitaldistribution at the birth county-sector-time period level or the birth county-sector-location-time period level (when location fixed effectsare included). Average initial capital is the mean of the distribution.Time period is an ordinal variable taking value from 1 to 3 corresponding to successive five-year time windows over the 1990-2004 period.Education and initial capital are obtained from the SAIC registration database and birth county population density is derived from the1982 population census.Standard errors clustered at birth county-sector level are reported in parentheses. * significant at 10%, ** at 5%, *** at 1%.
Conditional on having entered, the model predicts that firms from higher population density counties will
45The marginal entrant’s initial capital and the average entrant’s initial capital are now based on the distribution of capital ineach birth county-sector-destination-time period. The sample in Columns 4-5 is restricted to birth county-sector-destinations withentrants in the initial period. Similarly, the sample in Columns 2-3 is restricted to birth county-sectors with entrants in the initialperiod.
34
Figure 8. Asset Growth and Population Density
.1
.15
Ave
rage
ann
ual g
row
th o
f ass
ets
2004
-200
8
.025
.1
Ave
rage
ann
ual g
row
th o
f ass
ets
1995
-200
4
0 .02 .04 .06 .08 .1
Birth county population density
1995-2004 2004-2008
Time period:
Source: Industrial census (1995, 2004, 2008) and 1982 population census.Firm-level average annual growth of assets is averaged up to the birth county-sector level in each time period.
grow faster. Although the registration database is well suited to examine entry, concentration, and initial
capital investments, it is less suitable for analyses of capital growth. Registered capital does change, but given
that these changes are self-reported and involve substantial administrative costs, it will not track perfectly
with changes in the firm’s assets over time. For the analysis of firm growth, we thus turn (separately) to
the industrial census, which was conducted in 1995, 2004, and 2008 and the SAIC’s inspection database,
which includes annual firm-level information on assets and sales and which has reasonable coverage from 2004
onwards. We thus compute the average annual growth rate over the 1995-2004 and 2004-2008 periods with the
industrial census and, to be consistent, over the 2004-2008 period with the inspection data.46 Figure 8 reports
asset growth, separately in the 1995-2004 period and the 2004-2008 period, based on the industrial census.
The average annual growth of assets is increasing in population density in each time period and increasing
over time, as predicted by the model, in contrast with the patterns that we observe in the data for initial firm
size.
Table 11 reports parametric estimates corresponding to Figure 8. Since growth rates can only be computed
at two points in time with the industrial census data and the inspection data cover a relatively short period of
time, we focus on the cross-sectional predictions of the model. Note that the growth in firm size is computed
by differencing the level over time and, hence, the cross-sectional relationship between population density
and growth is effectively the change in the relationship between population density and the level of firm size
over time. To test the model’s predictions with industrial census data, we measure firm growth at the birth
county-sector level in Table 11, Columns 1 and 3 and at the birth county-sector-location level in Table 11,
Columns 2 and 4 and then estimate the (average) effect of birth county population density on these growth
measures. The estimating equation in Columns 1 and 3 includes sector fixed effects, while the estimating
46The average annual growth between period t and t′ is computed as the difference in log assets in t′ and t divided by t′ − t.Although there are no exits in the model, this is a feature of the data. In practice, firms with low profit levels – the young andthe less able – are more likely to exit. This selective exit, based on the profit level, does not bias our estimates because growthrates in the model are determined entirely by changes in CTFP that apply equally to all active firms in the network.
35
Table 11. Growth of Assets and Population Density
Dependent variable: average annual growth of assets
Source: industrial census industrial census inspection data
Time period: 1995-2004 2004-2008 2004-2008
(1) (2) (3) (4) (5) (6)
Birth county population density 0.006*** 0.007* 0.004** 0.003** 0.004*** 0.002*(0.002) (0.004) (0.002) (0.001) (0.001) (0.001)
Initial capital – 0.002*** – 0.001*** – 0.001***(0.000) (0.000) (0.000)
Mean of dependent variable 0.0528 0.0557 0.133 0.136 0.106 0.110Sector fixed effects Yes Yes Yes Yes Yes YesLocation fixed effect No Yes No Yes No YesObservations 5,517 5,664 31,234 64,258 18,701 43,470
Note: firm-level average annual growth of assets is averaged up to the birth county-sector level in specifications with sector fixed effectsand to the birth county-sector-location level in specifications with sector fixed effects and location fixed effects.Initial capital (in million Yuan) obtained from the SAIC registration database and birth county population density is derived from the1982 population census.Standard errors clustered at birth county level are reported in parentheses. * significant at 10%, ** at 5%, *** at 1%.
equation in Columns 2 and 4 includes sector fixed effects, location fixed effects, and the firm’s initial capital.
The fixed effects account for exogenous variation in firm growth across sectors and locations, which is not a
feature of our model. The firm’s initial capital is included to allow for convergence; recall that one alternative
explanation for why firms from high population density birth counties start small and then grow faster is
mechanical convergence (with initial size being accidentally determined). What we observe, instead, is that
firms that are larger to begin with, subsequently grow faster. Table 11, Columns 5-6 repeats the analysis
with SAIC inspection data, which include all sectors (not just manufacturing, as in the industrial census).47
The consistent finding across specifications is that population density in the birth county has a positive and
significant effect on firm growth.48
The estimating equation in Table 11 is specified to be consistent with the estimating equation in Table 10.
This allows for an appropriate comparison of initial size and firm growth. The main finding is that firms from
high population density birth counties start small but subsequently grow faster, after accounting for sector
and location effects. As discussed, this result is especially useful in distinguishing our model from alternative
non-network explanations.
5 Structural Estimation
Having validated the model, we next proceed to estimate its structural parameters. This will allow us to
quantify the contribution of the community networks to the growth in the number of firms and the capital
stock at the aggregate level. For the purpose of the structural estimation, we measure destinations by sectors
in order to restrict the dimensionality of the necessary computations. Two equations, with entry and average
initial capital in each birth county-sector-time period as the dependent variables, need to be estimated. Besides
47Data coverage for seven provinces is poor with the inspection data and these provinces are thus dropped from the analysis.48A pooled regression (not reported) which combines industrial census data over both time periods indicates, in addition, that
firm growth is increasing significantly over time. While our model can explain this result, it must be augmented, perhaps byintroducing credit constraints, to explain why initial capital has a positive and significant effect on firm growth.
36
the benchmark version of the model with myopic decision-making, we also estimate a variant (presented in
Appendix B) where entrepreneurs consider profits in the current period and the subsequent period when
making their entry decision.
Based on the corresponding equations in the myopic version of the model, (12) and (16), and retaining its
notation, we estimate the following structural equations:
eci,t = G(α, σ, r, A0)kcsci,t−1 +θ
(1− σ)(1− α)kcsci,t−1 · pnci,t−1 + uci,t (23)
logKaci,t = Jt(α, σ, r, A0, ft) +
θ(1− 2σ)
2(1− σ)(1− α)pnci,t−1 + vci,t (24)
eci,t measures the number of entrants and logKaci,t measures average initial capital (in logs) for birth county c
and sector i in time period t. We parameterize the θ(p) function to be increasing linearly in p: θ(p) = θp, with
the restriction that θ(0) = 0. The network effect is thus represented by a single parameter, θ. nci,t−1 is the
stock of firms from that birth county that are already established in that sector at the beginning of the time
period. sci,t−1 denotes the share of sector i in the stock of firms originating from county c at t− 1. Capital is
measured in the model in physical units, whereas in the data it is measured in monetary units. The mapping
from physical units to monetary units changes over time owing to changes in the price of capital goods. This
is especially relevant in the structural estimation because the objective is to match predicted and actual firm
size in each time period. ft thus represents the price of capital goods in period t.
kc measures the number of potential entrepreneurs from the birth county. The theoretical model assumed kc
was equal across birth counties and time periods. In practice, the number of potential entrepreneurs will depend
on the size of the population and the level of education in the county. The number of potential entrepreneurs
in each birth county is calculated from the 1990 population census, based on the characteristics of actual
entrepreneurs when they established their firms. We see in Appendix Figure D.4 that most entrepreneurs in
the SAIC database have at least high school education and that most were aged 25-44 when their firm was
established. kc is thus specified to be the number of men born in county c, aged 25-44, with at least high
school education, as reported in the 1990 population census.
The residual terms, uci,t and vci,t, measure the effect of local government inputs, agglomeration, and
sector-level spillovers on access to capital, firm productivity, and accompanying entry. We will see, with
an augmented specification, that adding sector-level spillovers to the structural equations does not affect the
estimated parameters, indicating that this component of uci,t, vci,t is uncorrelated with birth county population
density, p. Recall, however, from the tests of firm entry and firm size that entrepreneurs from higher p counties
appear to be selecting into less advantageous locations. The implied negative correlation between uci,t, vci,t
and p would bias the network spillover parameter, θ, downward and, hence, provide a conservative estimate
of the network effects.
The structural equations are linear in observed variables; (i) kcsci,t−1 (ii) kcsci,t−1 · pnci,t−1 (iii) pnci,t−1,
with four reduced-form coefficients.49 One of these coefficients, Jt, cannot be used to identify the structural
parameters because ft is unobserved. This leaves three reduced-form coefficients and five structural parameters:
49The functional forms for G(α, σ, r, A0) and Jt(α, σ, r, A0, ft) are obtained directly from (12) and (16), with the addition of theseparable ft term in the Jt function.
37
α, σ, r, A0, θ. We noted in Section 3 that the productivity channel and the credit channel for the network
effect cannot be separately identified. Although the model is parameterized to allow networks to increase
productivity, we remain agnostic about the specific channel through which the networks operate. For the
structural estimation, we specify that the network operates through the productivity channel as in the model,
setting r to 0.2 (which is in line with estimates of the average interest rate faced by Chinese firms).50 The
productivity multiplier is set to one in all sectors; i.e. A0 = 1. As in the model, variation in productivity
across sectors (and origin counties) in the benchmark specification is generated entirely by the network effect;
exp(θp · nci,t−1). In addition to the factors included above in the residual terms, we thus also abstract from
variation in product prices and labor productivity. The objective will be to assess how well our parsimonious
model is able to match the observed dynamics of entry and firm size.
To accommodate differences in the capital requirement across sectors, we do, however, allow the α param-
eter, which measures the marginal returns to capital, to vary across four broad sector categories: high-tech
services, wholesale and retail services, manufacturing and transportation, and heavy industry (mining, elec-
tricity, and construction). This increases the number of structural equations to eight, given that there are
now two equations in each sector category, and the number of structural parameters to be estimated to six;
α1, α2, α3, α4, σ, θ. The structural parameters are estimated by matching on entry and average initial capital
in each birth county-sector-time period.51 Initial entry in each birth county-sector is based on the number of
entrants in 1990-1994 and, if there is no entry in that time period, on the number of entrants in 1995-1999.
Sectors are defined at the one-digit level in the structural estimation to ensure that there is positive initial
entry in (almost) all birth county-sectors and the model is estimated over the 1995-2004 period; i.e. over two
time periods.
To estimate the structural parameters, we search for the set of parameters that minimize the distance
between the actual and the predicted entry and average initial capital; i.e. for which the sum of squared errors
over all birth county-sector-time periods is minimized. Parameter estimates, with bootstrapped standard errors
in parentheses, are reported for the benchmark model in Table 12, Column 1.52 The σ coefficient lies between
0.5 and 1, satisfying the condition, derived in the model, which ensures that average initial capital is decreasing
in birth county population density. The adjustment from physical capital to capital in monetary units, ft,
appears additively in the Jt function and, thus, can be estimated separately in 1995-1999 (period 1) and 2000-
2004 (period 2). Table 12, Column 2 reports estimates with the forward looking model, derived in Appendix
B. Entry must now be derived as the solution to a nonlinear equation, satisfying a fixed point condition, in
each birth county-sector-time period.53 Notice that the θ parameter declines substantially when we allow for
50In our model, r, is the sum of the real interest rate and the depreciation rate. Hsieh and Klenow (2009) assume that the realinterest rate is 0.05 in an economy, such as the U.S., with perfect financial markets and that the depreciation rate is 0.05. Usingthe same production function as Hsieh-Klenow and data from the Chinese industrial census, Brandt et al. (2016) estimate the realinterest rate to be 0.15 in 1995 and 2004 and 0.18 in 2008. We thus set r to 0.2.
51Although the number of reduced form coefficients now exceeds the number of structural parameters, the model places additionalrestrictions on the reduced form coefficients that must hold across sector categories. For example, the ratio of the coefficients onkcsci,t−1 · pnci,t−1 and pnci,t−1 in (23) and (24), respectively, must be 2/(1 − 2σ) in each sector category. The identification ofthe structural parameters is now more difficult to assess analytically and, hence, we verified that the parameters continue to be(just) identified by estimating the model with different values of r. The point estimates of the structural parameters do change inresponse, but the predicted entry and average initial capital (for each value of p) remain unchanged.
52When matching on entry and initial capital, we weight the error term by the reciprocal of the (bootstrapped) standarddeviation of the mean of each variable. The unweighted estimates are very similar to what we report in the table.
53The discount factor per year, δ is set to 0.8 when estimating the model with foresight. Because one time period is five years,
38
forward looking behavior, while remaining statistically and quantitatively significant. This is because potential
entrants require less of a “push” from the network when they take account of future benefits. Table 12, Column
3 reports estimates with an augmented model that allows for spillovers at the sector level, regardless of the
county of birth. This captures spillovers of the sort considered in endogenous growth models, resulting from
diffusion of R&D across firms in any given sector. An additional ni,t−1 term is thus included in equations (23)
and (24). Although the coefficient on ni,t−1, λ, is very precisely estimated, notice that the θ coefficient is very
similar in magnitude in Columns 1 and 3.
Table 12. Structural Estimates
Model: benchmark forward looking with sector-level spillovers
σ 0.775 0.776 0.777(0.001) (0.000) (0.001)
θ 0.395 0.125 0.359(0.054) (0.004) (0.011)
λ – – 0.00003(0.000002)
α1 0.137 0.119 0.138(0.009) (0.001) (0.007)
α2 0.189 0.179 0.194(0.020) (0.001) (0.020)
α3 0.192 0.177 0.197(0.011) (0.003) (0.016)
α4 0.236 0.226 0.240(0.004) (0.002) (0.004)
f1 0.223 0.233 0.216(0.004) (0.003) (0.016)
f2 0.189 0.208 0.184(0.005) (0.003) (0.012)
Note: the parameters are estimated by matching on entry and average initial capital (in logs), measured at the birth county-sector-timeperiod level. When minimizing the sum of squared errors, the error term is weighted by the reciprocal of the bootstrapped standarddeviation of the mean of each variable.The model is estimated over two time periods, 1995-1999 and 2000-2004, taking entry in the first period, 1990-1994, as given.Sectors are defined at the one-digit level when measuring entry and capital, but the α parameter is estimated at the aggregate sector level:(1) new technological services; (2) wholesale, retail and business service; (3) manufacturing and transportation; (4) heavy industry (mining,electricity, and construction)The discount factor per year δ is set to 0.8 in the forward looking model.Bootstrapped standard errors in parentheses.Number of entrants and average initial capital are computed from the SAIC registration database and birth county population density iscomputed from the 1982 population census.
Figure 9a assesses the goodness of fit of the benchmark model by comparing actual and predicted entry
across birth counties in each time period.54 The model that we estimate is extremely parsimonious, with just
six parameters. Nevertheless, it does a good job of predicting entry across nearly 2,000 birth counties and over
ten years. Figure 9b repeats this exercise with average initial capital and, once again, we see that the model
predicts variation across birth counties fairly accurately. Once the capital price adjustment factor is included
in each time period, note that actual and predicted average initial capital will match on levels by construction.
To formally test the goodness of fit of the model, we would want to compare predicted and actual outcomes
(entry and average initial capital) at each level of population density, p. Given the large number of counties,
with distinct values of p, we compare, instead, the estimated population density coefficient with actual and
predicted data (generated by each of the three models in Table 12). The dependent variable is either entry or
this works out to 0.82.5 = 0.56.54Birth county-sectors where initial entry commences in 1995-1999 are not included in the comparison for the 1995-1999 period.
39
Figure 9. Actual and Predicted, Entry and Initial Capital
0
1
2
3
4
Num
ber
of e
nter
ing
firm
s (t
hous
ands
)
0 .02 .04 .06 .08 .1
Birth county population density
Actual 1995-1999 Predicted 1995-1999
Actual 2000-2004 Predicted 2000-2004
(a) Entry
-.8
-.75
-.7
-.65
-.6
-.55
Ave
rage
initi
al c
apita
l (Lo
g)
0 .02 .04 .06 .08 .1
Birth county population density
Actual 1995-1999 Predicted 1995-1999
Actual 2000-2004 Predicted 2000-2004
(b) Initial Capital
Source: SAIC registration database, model generated data, and 1982 population census.
average initial capital.55 Since the structural estimates are based on two time periods, the estimating equation
includes birth county population density and a binary time period variable, which takes the value one in
2000-2004 and zero in 1995-1999. The estimated population density coefficients in Appendix Table E.5 are
statistically indistinguishable between actual data and data generated by the benchmark model. Moreover,
the benchmark model does a better job of matching the population density coefficient estimated with actual
data than both the forward looking model and the model that allows for additional sector-level spillovers.
An independent test of the structural model is to assess its out of sample predictions. The model is
estimated over the 1995-1999 and 2000-2004 periods. Figure 10 compares actual and predicted entry and
sectoral concentration in the subsequent 2005-2009 period.56 The model does a good job of predicting the
level of firm entry and sectoral concentration on average, although the predicted slope with respect to birth
county population density is higher than in the data.
With a single θ parameter common to all birth counties and sectors in the model, cross-sectional variation
is generated by differences in population density and initial entry alone. Nevertheless, the model is able to
match the data quite well across counties and sectors, even out of sample. The estimated parameters can
thus be used for counter-factual simulations. A major objective of our research is to quantify the role played
by community networks in the growth of private enterprise in China. This is accomplished by setting the θ
parameter to zero and then generating counter-factual entry and capital investment over the sample period.
The results of this exercise, with the benchmark model, are reported in Figure 11a. It is evident from the figure
that the number of entrants would have been substantially reduced in the absence of community networks,
particularly in higher population density birth counties. Based on our estimates, the total number of entrants
would have declined by 64% over the 1995-2004 period if the networks had not been active. In a related
counter-factual exercise, the total stock of capital in 2004 (taking account of the number of firms that entered,
55Entry is measured at the birth county-time period level and average initial capital is measured at the birth county-sector-timeperiod level to be consistent with the tests of the model above.
56We cannot test the model’s ability to predict initial capital beyond the sample period because the mapping from physicalcapital to capital in monetary units is unavailable.
40
Figure 10. Out of Sample Tests - Entry and Sectoral Concentration, 2005-2009
0
2
4
6
adju
sted
Her
finda
hl H
irsch
man
n In
dex
0
5
10
15
20
25
Num
ber
of e
ntra
nts
(tho
usan
ds)
0 .02 .04 .06 .08 .1
Birth county population density
Actual entrants Predicted entrants
Actual HHI Predicted HHI
Source: SAIC registration database, model generated data, and 1982 population census.Number of entrants and sectoral concentration computed at the birth county level in 2005-2009 with actual and model generated data.
their initial capital, and the subsequent growth in their capital) would have declined by 65% had the networks
been absent. Adjusting for the fact that our analysis is restricted to county-born entrepreneurs, for whom the
hometown networks are relevant, this amounts to a 40% decline in the number of entrants and the stock of
capital for the country as a whole.57
The preceding numbers are broadly in line with the results from a simple quantification exercise in which
we regress entry in each birth county-sector-location-time period on a full set of birth county dummies, location
dummies, and the interaction of these dummies with time. Although this exercise does not explicitly quantify
the birth county network effect, the advantage of the less structured approach is that we can account flexibly
for location effects. Based on this model, 45% of the predicted variation in entry over the 1995-2004 period
can be explained by variation across birth counties. In an independent robustness test, Figure 11b repeats
the counter-factual entry analysis with the augmented model that allows for sector-level spillovers. Notice
that these spillovers have almost no impact on entry, in contrast with the substantial impact that we estimate
for the birth county-sector spillovers. This indicates that origin-based networks constitute the main source of
spillovers in China, rather than the sector-based spillovers considered in the endogenous growth literature.
An important objective of industrial policy in any developing economy is to stimulate entrepreneurship.
It has been claimed that the government played a critical role in accelerating China’s growth by providing
firms with subsidized credit; e.g. Song et al. (2011), Wu (2016). In the absence of a market-based allocation
mechanism, a natural question to ask is which firms should have been targeted for the subsidy. To answer
this question, we examine a counter-factual policy experiment in which all entering firms in the 1995-1999
period received credit at an interest rate of 0.15; i.e. with a subsidy of 0.05. This subsidy would have had
two effects; it would have induced additional firms to enter at the margin and it would have increased the
57Government infrastructure and prices remain fixed in the counter-factual simulation. If the network were shut down and thenumber of firms declined, then output (input) prices would increase (decrease). The resulting increase in profits would encouragesome additional firms to enter. In contrast, if government infrastructure and the growth of the networks are complementary,then the removal of the networks would reduce the infrastructure level, generating a further decline in the number of firms in thecounter-factual scenario.
41
Figure 11. Counter-Factual Simulation: Effect of Community Networks on Entry
0
1
2
3
4
Num
ber
of e
nter
ing
firm
s (t
hous
ands
)
0 .02 .04 .06 .08 .1
Birth county population density
Predicted entry with networks Predicted entry without networks
(a) Benchmark quantification
0
1
2
3
Num
ber
of e
nter
ing
firm
s (t
hous
ands
)
0 .02 .04 .06 .08 .1
Birth county population density
Predicted entry without networksPredicted entry with networks
Predicted entry without networks and sector-level spillovers
(b) Quantification with sector-level spillovers
Source: Model generated data and 1982 population census.
profit of all (marginal and infra-marginal) entrants. As observed in Figure 12a, the total profit increase
generated by the subsidy in 1995-1999 is less than the cost to the government in all birth counties. However,
the spillover effect of the one-time subsidy on profits in the subsequent 2000-2004 period is substantial (and
even larger than the direct effect on profits in high population density counties). This is because the credit
subsidy induces additional entry during 1995-1999, which through the compounding network effect generates
large profit increases in the more socially connected counties in 2000-2004. With an annual discount factor
of 0.8, the return on the subsidy, based on the additional (discounted) profits that were generated over the
1995-1999 and 2000-2004 periods minus the cost of the subsidy, would have been 12% for countries above the
mean population density and -45% for counties below the mean.
Figure 12b reports the impact of an alternative government program, which only gives the subsidy to
those origin counties who would have increased their aggregate discounted profits over the 1995-2004 period
by more than the amount of subsidy they received in the preceding counter-factual experiment. To keep the
total amount of the subsidy constant, the interest rate for the targeted counties is now reduced to 0.11. The
increase in profits minus the subsidy received is reported across the population density distribution in the
figure, both for the original subsidy scheme and for the targeted subsidy scheme. As can be seen, the targeted
program does strictly better if the government’s objective is to maximize total profit (less the subsidy cost).
Notice also that average initial capital, which is declining with population density, declines even more steeply
with the more efficient targeted program.58 A distinguishing feature of our network-based mechanism is that
efficiency-enhancing policies could actually result in even smaller firms and even greater dispersion in firm
size in equilibrium (as observed in Figure 12b). The qualifier here is that this argument only applies if the
networks increase productivity. If the networks simply capture rents (cheap capital), and higher population
density birth counties have stronger networks, then a model based on community networks will still explain
58In our analysis, the marginal value product of capital does not vary across firms by construction. If we had used credit marketimperfections to motivate network formation instead, then a efficiency-enhancing policy that exploited network spillovers wouldhave increased the dispersion in marginal productivity within sectors (due to variation in interest rates across networks) as well.Haltiwanger et al. (2018) also make the point that an increase in efficiency could increase the dispersion in TFP, but with adifferent mechanism.
42
the stylized facts, but the normative interpretation will be very different. In particular, the preponderance of
small firms and wide variation in firm size will now be indicative of a misallocation. Recall, however, that we
found no evidence of rent seeking behavior, and the associated competition between community networks, in
our tests of the canonical model.
Figure 12. Counter-Factual Simulation: Effect of Interest Rate Subsidy on Profits
0
.5
1
1.5
Pro
fit/s
ubsi
dy (
billi
on Y
uan)
0 .02 .04 .06 .08 .1
Birth county population density
total subsidy(1995-1999) profit increase(1995-1999)
profit increase(2000-2004)
(a) Subsidy to all counties
-.9
-.8
-.7
-.6
-.5
Ave
rage
initi
al c
apita
l (Lo
g)
-1
-.5
0
.5
1
1.5
Pro
fit in
crea
se(b
illio
n Y
uan)
0 .02 .04 .06 .08 .1
Birth county population density
profit increase(1995-2004) minus subsidy: average initial capital:
subsidy to all counties subsidy to all counties
targeted subsidy targeted subsidy
(b) Targeted subsidy vs. subsidy to all counties
Source:Model generated data and 1982 population census.
6 Conclusion
In this paper, we identify and quantify the role played by community networks, organized around the birth
county, in the growth of private enterprise in China. The model that we develop generates predictions for the
dynamics of firm entry, sectoral and spatial concentration, and firm size across birth counties with different
levels of social connectedness (measured by population density) when networks are active. We validate each
of these predictions over a twenty year period with unique administrative data that covers the universe of
registered firms and provides information on entrepreneurs’ birth counties. The rich set of results that we
obtain, taken together, allow us to rule out alternative non-network based explanations. Additional results
provide direct support for the network channel, indicating that spillovers occur within the birth county and in
particular within clans within the county. Having validated the model, we estimate its structural parameters
and conduct counter-factual simulations. The first simulation indicates that aggregate entry and private
investment in China would have been 40% lower over the 1995-2004 period in the absence of the community
networks. While the contribution of these informal institutions to Chinese growth has thus been substantial,
in line with the anecdotal evidence, this still leaves room for other factors that have been associated with
this process, such as government policies, infrastructure, and finance; high saving rates and foreign investment
inflows; and the opening of the world market to Chinese exports.
The substantial inter-firm spillovers that we document are unlikely to be fully anticipated or internalized
by individual entrepreneurs. This creates scope for industrial policies to stimulate private investment, and
this is the subject of our second counter-factual simulation. This experiment, which simulates the effect of a
one-time credit subsidy, shows that the optimal strategy to maximize total profits would be to target entrants
43
from higher population density birth counties in order to take advantage of the larger resulting network
externalities. There are, however, a number of caveats to such a policy prescription. First, a policy that places
weight on both social affiliation and individual merit will only be effective in a population where community
networks are already active or have the potential to be activated, and this will depend on the underlying
social structure. In particular, the Chinese development experience will not be replicated in other countries
by simply providing infrastructure and credit. This is relevant for Chinese overseas development assistance
policy, which has largely focussed on infrastructure construction and industrial development (Zhang, 2016).59
Chinese development assistance has grown exponentially in recent years (Lin and Wang, 2016), but our analysis
indicates that the expected returns will only be realized if community networks in the recipient countries evolve
in parallel with the infrastructure construction, just as they did in China.
The second caveat concerns the normative consequences of such networks. Our analysis has primarily
focused on their positive or descriptive consequences. With regard to efficiency, credit subsidies targeted at
particular communities are justified if the resulting network spillovers increase firm productivity (and profits).
The potential downside is that communities will start to lobby the government for access to cheap capital; in
the extreme case, such rent seeking without accompanying productivity gains could worsen existing distortions
in the economy. Moreover, there are important consequences for inequality that need to be considered. By
bringing in less able entrepreneurs at the margin, community networks are redistributive within their popula-
tions. However, a policy that targets individuals from more socially connected populations to take advantage of
the positive externalities that their stronger networks provide will only exacerbate existing inequalities across
communities. Given the dynamic increasing returns generated by the networks, these inequalities will persist
and, if anything, worsen over time. Absent other redistributive mechanisms, any policy that attempts to
exploit network externalities must pay attention to the potentially enduring consequences for inter-community
inequality.
59This policy is explicitly motivated by the Chinese domestic experience, and the belief that infrastructure construction is thekey to development (see, for example, China’s second Africa policy paper; Xinhua, December 4, 2015).
44
References
Aghion, P., Cai, J., Dewatripont, M., Du, L., Harrison, A. and Legros, P. (2015). Industrial Policy
and Competition. American Economic Journal: Macroeconomics, 7 (4), 1–32.
— and Howitt, P. (1992). A Model of Growth through Creative Destruction. Econometrica, 60 (2), 323–351.
Akcigit, U., Alp, H. and Peters, M. (2016). Lack of Selection and Limits to Delegation: Firm Dynamics
in Developing Countries. NBER Working Paper Series.
Alesina, A. and Giuliano, P. (2015). Culture and institutions. Journal of Economic Literature, 53 (4),
898–944.
Allen, F., Qian, J. and Qian, M. (2005). Law, Finance, and Economic Growth in China. Journal of
Financial Economics, 77 (1), 57–116.
Asker, J., Collard-Wexler, A. and De Loecker, J. (2014). Dynamic Inputs and Resource (Mis) Allo-
cation. Journal of Political Economy, 122 (5), 1013–1063.
Au, C.-C. and Henderson, J. V. (2006). Are Chinese Cities Too Small? Review of Economic Studies, 73 (3),
549–576.
Bai, C.-e., Hsieh, C.-T. and Song, Z. (2019). Special deals with chinese characteristics.
Banerjee, A. and Munshi, K. (2004). How Efficiently is Capital Allocated? Evidence from the Knitted
Garment Industry in Tirupur. Review of Economic Studies, 71 (1), 19–42.
Brandt, L., Kambourov, G. and Storesletten, K. (2016). Firm Entry and Regional Growth Disparities:
the Effect of SOEs in China. University of Toronto Mimeo.
—, Van Biesebroeck, J. and Zhang, Y. (2012). Creative Accounting or Creative Destruction? Firm-level
Productivity Growth in Chinese Manufacturing. Journal of Development Economics, 97 (2), 339–351.
Brooks, W. J., Kaboski, J. P. and Li, Y. A. (2016). Growth Policy, Agglomeration, and (the Lack of)
Competition. NBER Working Paper Series.
Caunedo, J. (2016). Industry Dynamics, Investment and Uncertainty. Working Paper.
Ciccone, A. and Hall, R. E. (1996). Productivity and the Density of Economic Activity. American Eco-
nomic Review, 86 (1), 54–70.
Coleman, J. S. (1988). Social Capital in the Creation of Human Capital. American Journal of Sociology, 94,
S95–S120.
Combes, P.-P., Duranton, G., Gobillon, L., Puga, D. and Roux, S. (2012). The Productivity Advan-
tages of Large Cities: Distinguishing Agglomeration from Firm Selection. Econometrica, 80 (6), 2543–2594.
Duranton, G. and Overman, H. G. (2005). Testing for Localization Using Micro-Geographic Data. Review
of Economic Studies, 72 (4), 1077–1106.
Ellison, G. (1994). Cooperation in the Prisoner’s Dilemma with Anonymous Random Matching. Review of
Economic Studies, 61 (3), 567–588.
— and Glaeser, E. L. (1997). Geographic Concentration in US Manufacturing Industries: a Dartboard
Approach. Journal of Political Economy, 105 (5), 889–927.
Fang, C. (1997). Recent Trends of Migration and Population in Shandong. Working Papers on Chinese
Politics, Economy and Society, (10).
45
Fisman, R., Shi, J., Wang, Y. and Xu, R. (2018). Social Ties and Favoritism in Chinese Science. Journal
of Political Economy, 126 (3), 1134–1171.
Fleisher, B., Hu, D., McGuire, W. and Zhang, X. (2010). The Evolution of an Industrial Cluster in
China. China Economic Review, 21 (3), 456–469.
Goodman, B. (1995). Native Place, City, and Nation: Regional Networks and Identities in Shanghai,
1853–1937. University of California Press.
Greif, A. (1993). Contract enforceability and economic institutions in early trade: The maghribi traders’
coalition. The American economic review, pp. 525–548.
— (1994). Cultural beliefs and the organization of society: A historical and theoretical reflection on collectivist
and individualist societies. Journal of political economy, 102 (5), 912–950.
— and Tabellini, G. (2017). The Clan and the Corporation: Sustaining Cooperation in China and Europe.
Journal of Comparative Economics, 45 (1), 1–35.
Haltiwanger, J., Kulick, R. and Syverson, C. (2018). Misallocation measures: The distortion that ate
the residual. Tech. rep., National Bureau of Economic Research.
Honig, E. (1992). Creating Chinese Ethnicity: Subei People in Shanghai, 1850-1980. New Haven: Yale
University Press.
— (1996). Regional Identity, Labor, and Ethnicity in Contemporary China. Berkeley, California: Institute of
East Asian Studies, University of California.
Hsieh, C.-T. and Klenow, P. J. (2009). Misallocation and Manufacturing TFP in China and India. Quarterly
Journal of Economics, 124 (4), 1403–1448.
— and — (2014). The Life Cycle of Plants in India and Mexico. Quarterly Journal of Economics, 129 (3),
1035–1084.
Hu, B. (2008). People’s Mobility and Guanxi Networks: A Case Study. China & World Economy, 16 (5),
103–117.
Huang, Z., Zhang, X. and Zhu, Y. (2008). The Role of Clustering in Rural Industrialization: A Case Study
of the Footwear Industry in Wenzhou. China Economic Review, 19 (3), 409–420.
Jackson, M. O., Rodriguez-Barraquer, T. and Tan, X. (2012). Social capital and social quilts: Network
patterns of favor exchange. American Economic Review, 102 (5), 1857–97.
— and Rogers, B. W. (2007). Meeting strangers and friends of friends: How random are social networks?
American Economic Review, 97 (3), 890–915.
Jones, C. I. (1995). R & d-Based Models of Economic Growth. Journal of Political Economy, 103 (4),
759–784.
Kandori, M. (1992). Social Norms and Community Enforcement. Review of Economic Studies, 59 (1), 63–80.
Levine, R. and Rubinstein, Y. (2017). Smart and Illicit: Who Becomes an Entrepreneur and Do They Earn
More? Quarterly Journal of Economics, 132 (2), 963–1018.
Lin, J. Y. and Wang, Y. (2016). Going Beyond Aid: Development Cooperation for Structural Transformation.
Cambridge University Press.
Long, C. and Zhang, X. (2011). Cluster-based Industrialization in China: Financing and Performance.
Journal of International Economics, 84 (1), 112–123.
46
Ma, L. J. C. and Xiang, B. (1998). Native Place, Migration and the Emergence of Peasant Enclaves in
Beijing. The China Quarterly, 155, 546.
Morten, M. (2019). Temporary migration and endogenous risk sharing in village india. Journal of Political
Economy, 127 (1), 000–000.
Munshi, K. (2011). Strength in Numbers: Networks as a Solution to Occupational Traps. Review of Economic
Studies, 78 (3), 1069–1101.
— (2014). Community Networks and the Process of Development. Journal of Economic Perspectives, 28 (4),
49–76.
— and Rosenzweig, M. (2006). Traditional Institutions Meet the Modern World: Caste, Gender, and
Schooling Choice in a Globalizing Economy. American Economic Review, 96 (4), 1225–1252.
— and — (2016). Networks and Misallocation: Insurance, Migration, and the Rural-Urban Wage Gap. Amer-
ican Economic Review, 106 (1), 46–98.
Nee, V. and Opper, S. (2012). Capitalism from Below: Markets and Institutional Change in China. Harvard
University Press.
Peng, Y. (2004). Kinship Networks and Entrepreneurs in China’s Transitional Economy. American Journal
of Sociology, 109 (5), 1045–1074.
Peters, M. (2016). Heterogeneous Markups, Growth and Endogenous Misallocation. Working Paper.
Restuccia, D. and Rogerson, R. (2008). Policy Distortions and Aggregate Productivity with Heterogeneous
Establishments. Review of Economic Dynamics, 11 (4), 707–720.
Romer, P. M. (1986). Increasing Returns and Long-run Growth. Journal of Political Economy, 94 (5),
1002–1037.
— (1990). Endogenous Technological Change. Journal of Political Economy, 98 (5, Part 2), S71–S102.
Ruan, J. and Zhang, X. (2009). Finance and Cluster-based Industrial Development in China. Economic
Development and Cultural Change, 58 (1), 143–164.
Segerstrom, P. S. (1998). Endogenous Growth without Scale Effects. American Economic Review, pp.
1290–1310.
—, Anant, T. C. and Dinopoulos, E. (1990). A Schumpeterian Model of the Product Life Cycle. American
Economic Review, pp. 1077–1091.
Song, Z., Storesletten, K. and Zilibotti, F. (2011). Growing Like China. American Economic Review,
101 (1), 196–233.
Summers, L. (2007). The Rise of Asia and the Global Economy. Research Monitor, (2).
Tsai, L. L. (2007). Solidary Groups, Informal Accountability, and Local Public Goods Provision in Rural
China. The American Political Science Review, 101 (2), 355–372.
Wu, M. (2016). The China, Inc. Challenge to Global Trade Governance. Harv. Int’l LJ, 57, 261.
Zhang, C. (2017). Culture and the Economy: Clan, Entrepreneurship, and Development of the Private Sector
in China. SSRN Scholarly Paper, (ID 2865105).
— and Xie, Y. (2013). Place of Origin and Labour Market Outcomes Among Migrant Workers in Urban
China. Urban Studies, 50 (14), 3011–3026.
Zhang, J. (2016). Chinese Foreign Assistance, Explained. Retrieved from
47
https://www.brookings.edu/blog/order-from-chaos/2016/07/19/chinese-foreign-assistance-explained/.
Zhang, X. and Li, G. (2003). Does Guanxi Matter to Nonfarm Employment? Journal of Comparative
Economics, 31 (2), 315–331.
Zhao, Y. (2003). The Role of Migrant Networks in Labor Migration: the Case of China. Contemporary
Economic Policy, 21 (4), 500–511.
48
Appendix A. Derivation of the Adjusted HHI
Suppose that there are n trials, that each outcome j from the set of k possible outcomes has an independent
probability of occurring pj , and that the random variable Xj is the number of occurrences of outcome j.
Then the multivariate random variable X = (X1, · · · , Xk) has a multinomial distribution with parameters
(n, k, p1, · · · , pk). Applied to our context, (i) n is the total number of entrepreneurs or firms from a given
birth county, (ii) k is the total number of clans or destinations that they are allocated to, and (iii) p1, · · · , pkare the probabilities that an entrepreneur or firm allocated randomly would end up in each of those clans or
destinations. We assume that there is an equal probability of choosing any clan or destination; pj = 1k , ∀j.
The expected HHI when firms make decisions independently can be expressed as,
E(HHI) = E
(1
n2
k∑i=1
X2i
)= E
(1
n2XTX
).
Based on the general properties of the multinomial distribution,
E(HHI) =1
n2
([E(X)]TE(X) + tr[cov(X)]
).
It follows that,
E(HHI) =1
n2
(k(nk
)2+ k
[n
1
k
(1− 1
k
)])=
1
k+
1
n
k − 1
k.
For large n, E(HHI) ≈ 1k . For small n, E(HHI) is decreasing in n. We account for this by constructing
a normalized HHI statistic, which is simply the unadjusted HHI, based on the observed distribution of en-
trepreneurs or firms across clans or destinations, divided by E(HHI). If entrepreneurs or firms are allocated
randomly, then the adjusted HHI will be close to one, providing a useful benchmark for this statistic.
Appendix B. Entry with Foresight
Consider the consequences of allowing entrepreneurs to look ahead and incorporate profits they would expect
to make after the first period they enter. We suppose cohort t agents look ahead one additional period, i.e.,
make their entry decision based on anticipated present value profits in periods t and t + 1. The equilibrium
can no longer be computed recursively, owing to the need for entrants to coordinate their expectations of entry
decisions of one another. We shall consider equilibria where these expectations are fulfilled. We continue to
assume that incumbents are committed to their previous entry decisions.
Let ξ denote ψr−α
1−α , and δ ∈ (0, 1) denote the common discount factor of agents. Then expected present
value of entering Bi at t for a cohort t agent of ability ω is
Pit(ω) = ωξA1
1−α0 exp(θpni,t−1
1
1− α)[1 + δ exp(θ(p)eit
1
1− α)] (25)
while of staying in T is
Nit(ω) = ωσ[1 + δ] (26)
The agent will enter if
logω >1
1− σ[− log ξ − 1
1− αlogA0 + log(1 + δ)− 1
1− αθ(p)ni,t−1 − log{1 + δ exp(θ(p)eit
1
1− α)}] (27)
49
Define the function
g(e|si,t−1, ni,t−1, Ai0) = ksi,t−1{1 + 11−σ [log ξ + 1
1−α logA0 − log(1 + δ)
+ 11−αθ(p)ni,t−1 + log{1 + δ exp(θ(p)e 1
1−α)}]}
Then equilibrium entry decisions form a fixed point of this function, i.e., eit = e(si,t−1, ni,t−1, Ai0) solves
g(e|si,t−1, ni,t−1, Ai0) = e (28)
The intercept of this function is exactly the entry that results in the myopic equilbrium with δ = 0. The
function is increasing in e, with a slope
g′(e|si,t−1, ni,t−1, Ai0) = si,t−1
δ exp( θ(p)e1−α )
1 + δ exp( θ(p)e1−α )
kθ(p)
(1− α)(1− σ)(29)
Hence ifkθp̄
(1− α)(1− σ)< 1 (30)
where p̄ is an upper bound for p, an equilibrium exists and is unique. Computing the equilibrium is easy, as
it involves solving for fixed points of a contracting mapping defined recursively by past entry decisions. It can
be easily verified that entry is rising in si,t−1, θ(p) and ni,t−1, just as in the myopic entry case.
Appendix C: Proofs
Proof of Proposition 1: We first prove that sit > si−1,t for all t. Suppose this is true at t − 1: then si,t−1 is
rising in i. Denote the growth rate of destination i share: git ≡ sit−si,t−1
si,t−1= Nt−1
Nt+ [L + κ(p)Nt−1]si,t−1 − 1,
upon using (12). Hence git is rising in si,t−1 and therefore in i, implying sit > si−1,t. So shares are ordered for
all cohorts exactly as they are in cohort 0. Also note that all destinations have positive shares in all cohorts,
and growth rates cannot be zero at any t for all destinations.
Since Ht ≡∑
i s2it =
∑i s
2i,t−1(1 + git)
2 =∑
i s2i,t−1 +
∑i s
2i,t−1g
2it + 2
∑i s
2i,t−1git, it follows that
Ht −Ht−1 =∑i
s2i,t−1g
2it + 2
∑i
s2i,t−1git
> 2∑i
s2i,t−1git
> 0
where the first inequality follows from the fact that all sector shares are positive and growth rates are not
all zero. The second inequality follows from observing that: (i) if we define xit ≡ si,t−1git = sit − si,t−1 then∑i xit = 0; (ii)
∑i si,t−1 = 1, and (iii) xit and si,t−1 are both increasing in i, as explained above. Hence by a
standard argument60∑
i s2i,t−1git =
∑i si,t−1xit > 0, which proves that concentration is rising in t, and hence
(using (13)) the same is true for Et.
60The distribution across destinations first order stochastically dominates the uniform distribution, in which si,t−1 is the samefor all i, and the expected value of x under the uniform distribution equals zero. Hence the expected value of x must be positive.
50
Proof of Proposition 2: The increase in Et, Nt, Ht with t follows from Proposition 1. So consider how a higher
p alters the dynamics, given initial conditions. We claim that it raises aggregate entry Et (and hence Nt) as
well as Ht at every date t. This follows from an inductive argument. Observe first that it must be true for
Et (and Nt) at t = 1, given the initial conditions N0, H0, upon applying equation (13) at t = 1. Next observe
that the right-hand-side of (14) is rising in p, given any Nt−1 and s1,t−1 >12 . Hence s11 must be rising in p,
given the initial conditions. So the result holds for Ht at t = 1. Next suppose it holds until some date t− 1,
i.e., Nt−1 and Ht−1 are rising in p. Equation (13) then implies Et (and Nt) is rising in p. Moreover, observe
that the right-hand-side of (14) is rising in Nt−1 and in s1,t−1, given p and s1,t−1 >12 . The share s1t will then
be increasing in p because it is increasing in s1,t−1, Nt−1 and κ(p) respectively. Induction now ensures this
will be true at every t. This establishes part (a) of Proposition 2.
Turn now to part (b). Taking first differences of (13)
Et+1 − Et = κ(p)[NtHt −Nt−1Ht−1] = κ(p)[EtHt +Nt−1(Ht −Ht−1)] (31)
Since κ,Et, Ht, Nt−1 are all rising in p, the result would hold for entry if it were also true for concentration
(i.e., Ht−Ht−1 is rising in p). A sufficient condition for this to hold is that it is true for s1t: i.e., if s1,t− s1,t−1
is rising in p (since Ht−Ht−1 = 2(s1t− s1,t−1)(s1t + s1,t−1− 1), and we have already shown that s1t, s1,t−1 are
rising in p).
Now observe that (14) can be rewritten as
s1t − s1,t−1 = κ(p)Nt−1(2s1,t−1 − 1)(1− s1,t−1)s1,t−1
(L+Nt−1)(2− s1,t−1) + κ(p)Nt−1(s21,t−1 + 1− s1,t−1)
(32)
κ(p) < 1 implies that the denominator of the right-hand-side of (32) is decreasing in s1,t−1. And the numerator
is increasing in s1,t−1 if s1,t−1 <34 (since this implies s1,t−1(1 − s1,t−1) > 1
6). Then s1t − s1,t−1 is rising in
s1,t−1, as well as in Nt−1 and κ. Part (b) then follows from the fact that s1,t−1, Nt−1 are rising in p.
Proof of Proposition 3: To verify (a), observe that averaging (15) across destinations (and noting that∑i si,t−1 = 1), initial capital of the marginal entrant is decreasing in t, p, and p ∗ t because θ(p) is in-
creasing in p and Nt−1 is increasing in t and p (more steeply over time).61 A similar argument operates for
ability and size of the average entrant from (16) and taking the average across destinations. Part (b) follows
from averaging across destinations in (18), and applying Propositions 1 and 2.
61Nt is increasing in t (for any given p) and in p (for any given t) from Propositions 1 and 2. Nt−Nt−1 ≡ Et, which is increasingin p from Proposition 2, hence, the cross-partial derivative of Nt with respect to p and t is positive.
51
Appendix D. Figures
Figure D.1. Number of Firms, By Type
0
5
10
15
Num
ber
of fi
rms
(mill
ion)
1990 1995 2000 2005 2010 2015
year
Private SOE TVE Foreign
Source: SAIC Registration Database.
Figure D.2. Population Density over Time
0
.05
.1
.15
Popu
latio
n de
nsity
eac
h ye
ar
0 .02 .04 .06 .08 .1
1982 population density
1982 1990 2000 2010
Source: Registration Database and 1982 population census.
52
Figure D.3. Firm Entry and Population Density
0
2
4
6
8
Num
ber
of fi
rms
in s
tock
(th
ousa
nds)
0 .02 .04 .06 .08 .1
Birth county population density
1994 1999 2004 2009
Year:
(a) Stock of Firms
0
.5
1
1.5
2
2.5
Num
ber
of e
nter
ing
firm
s (t
hous
ands
)
0 .02 .04 .06 .08 .1
Birth county population density
1990-1994 1995-1999 2000-2004 2005-2009
Time period:
(b) Firms Located Outside the Birth County
Source: Registration Database and 1982 population census.
Figure D.4. Education and Age Distribution of Entrepreneurs
0
.01
.02
.03
.04
.05
Fra
ctio
n
20 25 30 35 40 45 50 55 60 65 70
Age when starting firms
Actual Potential entrepreneur
(a) Age
0
.1
.2
.3
.4
.5
Fra
ctio
n
Primary Middle High Junior Colledge Undergraducate
Education
Actual Potential entrepreneur
(b) Education
Source: SAIC registration database.
53
Appendix E. Tables
Table E.1. Composition of Listed Individuals in the Firm, by Birth Place Population Density
Dependent variable: fraction of listed individuals from the legal rep-resentative’s birth county
Birth place: couty city
(1) (2)
Birth place population density 0.010*** -0.017***(0.002) (0.001)
Mean fraction 0.741 0.606Counter-factual fraction with random assignment 0.0560 0.0916Observations 490,273 245,302
Source: State Administration of Industry and Commerce (SAIC) registration database.Sample restricted to firm’s operating outside their legal representative’s birth county.Counter-factual fraction with random assignment for such a firm is measured by the number of listed individuals from the legal representa-tive’s birth county in its sector-location divided by the total number of listed individuals (among all firms whose legal representatives wereborn outside their birth county) in that sector-location. The sector is measured at the 2-digit level and location is the county or urbandistrict.“Listed individuals” includes major investors and top managers, but excludes the legal representative.Population density in the county or city based on the 1982 population census.Population density is measured in units of 10,000 people per square km, and then converted to Z-score.Standard errors clustered at the county or city level reported in parentheses. * significant at 10%, ** at 5%, *** at 1%.
Table E.2. The Effect of Initial Entry on Subsequent Entry (for firms located outside the birth county)
Dependent variable: subsequent entrants from the birth place
Birth place: county county city
Time period: 2000-2004 2005-2009 2000-2004 2005-2009 2000-2004 2005-2009
(1) (2) (3) (4) (5) (6)
Initial entrants from the birth place 5.481*** 8.273*** 3.303*** 5.883** 6.813*** 5.152***(1.044) (1.948) (1.197) (2.470) (0.506) (0.507)
All initial entrants at the destination 0.294*** 0.442*** – – – –(0.036) (0.047)
Initial entrants from the birth place ×birth place population density
– – 1.820*** 1.880* 0.081 -0.083
(0.507) (0.966) (0.087) (0.080)Distance to the birth place – – -1.458*** -1.723*** -2.831*** -2.609***
(0.077) (0.069) (0.155) (0.101)
Mean of dependent variable 1.964 2.124 1.964 2.124 3.382 2.926Origin-sector fixed effects Yes Yes Yes Yes Yes YesLocation fixed effects No No Yes Yes Yes YesObservations 334,843 713,935 334,843 713,935 289,708 420,138Note: number of entrants outside the birth place is measured at the (two-digit) sector-destination level.Initial entry is derived over the 1990-1994 period.Number of entrants is obtained from the SAIC registration database and birth county population density is derived from the 1982 populationcensus.Standard errors clustered at birth county-sector level are reported in parentheses. * significant at 10%, ** at 5%, *** at 1%.
54
Table E.3. Clan Concentration and Population Density (time and interaction effects)
Dependent variable: clan concentration
Sample: all firms firms located outside thebirth county
(1) (2)
Time period 1352*** 1.027***(0.014) (0.012)
Birth county population density × time period 0.254*** 0.238***(0.015) (0.014)
Observations 6,482 6,300
Note: the estimating equation includes, in addition, birthplace population density and a constant term.Clan is defined by the same surname and the same birth county.Clan concentration measured by Herfindahl Hirschman Index (HHI) across clans divided by the expected HHI that would be obtained byrandom assignment.Population density is measured in units of 10,000 people per square km, and then converted to Z-score.Time period is an ordinal variable taking value from 1 to 4 corresponding to successive five-year time windows over the 1990-2009 period.Clan concentration is derived from the SAIC registration database and population density is derived from the 1982 population census.Standard errors clustered at birth county level are reported in parentheses. * significant at 10%, ** at 5%, *** at 1%.
Table E.4. Marginal Ability, Marginal Initial Capital and Population Density
Dependent variable: marginal ability marginal initial capital
Time period: 1990-1994 1995-1999 2000-2004 2005-2009 1990-1994 1995-1999 2000-2004 2005-2009
(1) (2) (3) (4) (5) (6) (7) (8)
Birth county populationdensity
-1.829 -2.685*** -3.570*** -1.435*** -0.100*** -0.118*** -0.134*** -0.041***
(1.369) (1.103) (0.807) (0.755) (0.030) (0.026) (0.017) (0.010)
Mean of dependent variable 66.07 52,47 40.79 40.15 -0.803 -1.093 -1.669 -2.225Observations 4,079 6,595 10,354 11,137 15,601 46,877 83,276 99,877
Note: the entrepreneur’s ability is measured by his percentile rank in his birth county-birth cohort education distribution.The marginal entrepreneur is defined as the individual at the bottom one percentile of the ability distribution among entering entrepreneursin a given birth county-sector-time period. Marginal initial capital defined as the bottom one percentile of the initial capital distributionat the birth county-sector-time period level.Control variables include population, education and occupation distribution in the birth county.Standard errors clustered at birth county level are reported in parentheses. * significant at 10%, ** at 5%, *** at 1%.
Table E.5. Estimates based on Actual and Model Generated Data
Dependent variable: number of entrants average initial capital
Data: actualmodel
generated(myopic)
modelgenerated(forwardlooking)
modelgenerated
(withsector-levelspillovers)
actualmodel
generated(myopic)
modelgenerated(forwardlooking)
modelgenerated
(withsector-levelspillovers)
(1) (2) (3) (4) (5) (6) (7) (8)
Birth countypopulation density 0.335*** 0.215*** 0.109*** 0.171*** -0.010** -0.002** -0.003** -0.003**
(0.034) (0.030) (0.016) (0.025) (0.005) (0.001) (0.001) (0.001)Time period 0.476*** 0.305*** 0.095*** 0.369*** -0.202*** -0.191*** -0.151*** -0.191***
(0.020) (0.021) (0.011) (0.019) (0.009) (0.002) (0.003) (0.002)
Observations 3,032 3,032 3,032 3,032 18,401 18,401 18,401 18,401
Note: Estimates based on two time periods, 1995-1999 and 2000-2004.Firm entry (in thousands) measured at the birth county-time period level. Initial capital (in million Yuan) is measured in logs. Averageinitial capital is the average of the initial capital distribution at the birth county-sector-time period level.Time period is a binary variable taking the value one for the 2000-2004 period and zero for the 1995-1999 period.Actual firm entry and initial capital are obtained from the SAIC registration database and birth county population density is derived fromthe 1982 population census.Standard errors clustered at the birth county level are reported in parentheses. * significant at 10%, ** at 5%, *** at 1%.
55